TL;DR: AI knowledge management combines retrieval augmented generation, enterprise search, knowledge graphs, and AI agents to turn scattered documents and tribal knowledge into governed answers with citations, delivered inside the tools employees already use. Start with one high-friction domain, wire permissions into retrieval from day one, and measure time to answer and ticket deflection before you scale.
Watch on YouTube
DokGPT | AI-Powered RAG Chatbot for Smart Document Search
See a retrieval augmented generation chatbot answer questions over real enterprise documents, the working example of the architecture this article walks through.
Ask a simple question in most enterprises and the answer already exists somewhere. It sits in a wiki page nobody updated, a contract buried in SharePoint, or the head of an engineer who resigned in March.
Generative AI has reset expectations for how fast that answer should arrive. Employees who get instant, sourced responses from consumer chat tools now wonder why finding the company travel policy takes forty minutes and three Slack messages.
The distance between what an organization knows and what its people can retrieve has become a measurable cost, and closing it is now a board-level program rather than an intranet cleanup. In this article, we’ll cover what AI knowledge management is, the architecture behind it, use cases by function, build versus buy, a 90-day roadmap, and how to measure results.
Why Tribal Knowledge Became the Enterprise Bottleneck Deloitte’s Global Human Capital Trends research found that 75 percent of organizations rate creating and preserving knowledge as important or very important to their success, yet only 9 percent feel very ready to address it. That readiness gap has a price. Every resignation takes context with it, every reorganization orphans a set of documents, and every new hire spends weeks rediscovering decisions that were already made.
The pressure is rising as AI adoption accelerates. Stanford’s AI Index reports that 78 percent of organizations used AI in 2024, up from 55 percent the year before. Most of those deployments reason over public training data, while the answers employees actually need live in the company’s own contracts, runbooks, tickets, and inboxes.
Tribal knowledge, the unwritten know-how that lives in people’s heads, is the most expensive kind to lose. AI knowledge management exists to capture it, connect it to documented sources, and serve it back on demand to the next person who asks.
Key Takeaways AI knowledge management uses natural language processing, retrieval augmented generation, and AI agents to convert scattered enterprise content into direct, cited answers. Traditional knowledge bases return documents to read, while AI systems synthesize answers, respect permissions per query, and keep improving as content changes. Production systems follow a five-layer architecture, from knowledge sources through ingestion, retrieval, and LLM reasoning to serving with feedback. Support, engineering, sales, HR, and compliance teams see the fastest payback because they answer the same questions repeatedly under time pressure. Most enterprises land on a hybrid approach, buying retrieval infrastructure while owning their data pipelines, permissions, and evaluation sets. Kanerika’s DokGPT deployments show what governed retrieval delivers, including 43 percent faster information retrieval at an investment bank with full role-based access control. What Is AI Knowledge Management? AI knowledge management is the use of artificial intelligence, including natural language processing , large language models, and machine learning, to capture, organize, retrieve, and apply an organization’s collective knowledge. Instead of asking employees to hunt through folders, the system understands a question in plain language and assembles an answer from every source it has permission to see.
The discipline itself is decades old. What changed is the interface and the reach. A knowledge system can now read a messy contract, summarize a two-hour meeting, and answer a follow-up question in the same conversation, which turns knowledge management from a filing exercise into a working tool.
Listen on Spotify
The Smarter Way to Access Enterprise Knowledge – With DokGPT
IBM’s framing of the field distinguishes three types of knowledge that any program has to handle.
Explicit knowledge. Documented content such as policies, contracts, specifications, and reports. Easy to store, easy to search badly.Implicit knowledge. Applied know-how transferred through practice, such as how a team actually runs a production deployment versus what the runbook says.Tacit knowledge. Personal expertise and judgment that resists documentation, like knowing which stakeholder really makes the renewal decision.Earlier generations of knowledge software handled only the first category well. AI extends coverage to the other two by mining tickets, chat threads, and call transcripts, the same pattern behind many generative AI use cases , and by making contribution cheap enough that experts document a decision once instead of repeating it in meetings for a year.
How AI Knowledge Management Differs from a Traditional Knowledge Base A traditional knowledge base is a library. It stores documents, arranges them into categories, and depends on keyword search plus human curation to stay usable. The model works until the library grows past what curators can maintain, which in most enterprises happened years ago.
An AI knowledge system behaves less like a library and more like a well-briefed colleague. It reads the question for intent, retrieves passages rather than files, and returns a direct answer with citations so the reader can verify the source. The difference between retrieval-grounded answers and raw model output is worth understanding, and our RAG vs LLM comparison walks through it.
Dimension Traditional Knowledge Base AI Knowledge Management Search Keyword match on titles and tags Semantic search that understands intent and synonyms Answer format A list of documents to read A direct answer with source citations Maintenance Manual tagging and periodic audits Automatic classification with freshness and duplicate signals Tacit knowledge Lost unless someone writes it down Mined from tickets, chats, and meeting transcripts Permissions Folder-level access rules Permission-aware retrieval enforced on every query Adoption curve Peaks at launch, decays as content goes stale Compounds as usage feeds evaluation and coverage
Table 1: Traditional knowledge base vs AI knowledge management across the dimensions that decide adoption.
The last row matters most for anyone who has watched a wiki die. Static repositories decay because maintenance is a chore separate from the work. AI systems improve with use because every query, rating, and correction becomes training signal for retrieval quality.
Six Jobs AI Takes Over in the Knowledge Lifecycle Search gets the attention, but the deeper change is in who does the maintenance work. Across the capture, organize, deliver, and maintain cycle, AI absorbs the chores that made knowledge management fail as a human discipline.
Automatic capture. Meeting transcripts, resolved tickets, and project channels become draft knowledge articles without anyone opening an editor. The expert reviews instead of writes, which drops the contribution cost by an order of magnitude.Classification and tagging. Models assign topics, products, audiences, and sensitivity labels as content arrives, replacing the manual taxonomy work that curators abandoned first.Summarization. A 60-page contract becomes a one-paragraph brief with links back to the exact clauses, so readers choose their own depth instead of skipping the document.Expert location. When no document answers the question, the system identifies who has answered similar ones before, turning the org chart into a searchable map of tacit knowledge.Freshness and gap detection. Usage analytics flag articles that contradict newer sources, pages nobody reads, and questions that keep failing, which converts maintenance from a quarterly audit into a ranked queue.Answer synthesis. The visible payoff, a direct response assembled from several sources with citations, delivered in the channel where the question was asked.Together these six shift the human role from librarian to editor. People still decide what is true, and the system handles everything about getting that truth to the next person who needs it.
The Core Components of an AI Knowledge Management Architecture Vendor labels differ, but production systems resolve to the same five-layer pattern. Understanding it keeps procurement conversations honest and makes the build versus buy decision tractable. Procurement teams applying the same retrieval pattern to supplier and commodity data will find a specialized analysis in the AI for procurement guide. Building guardrails into the retrieval layer early—access controls, audit trails on model queries, and output guardrails—maps to the framework in AI governance best practices .
Knowledge sources. Wikis, document stores, email, CRM notes, code repositories, support tickets, and meeting transcripts, kept in place rather than copied into yet another silo.Ingestion and indexing. Pipelines that clean, chunk, and embed content, attach metadata and access controls, and re-index on a schedule so answers track reality.Retrieval. Hybrid search that blends vector similarity with keyword precision and metadata filters, then reranks candidates before anything reaches a model.Reasoning. Large language models generate grounded answers through retrieval augmented generation , while agents plan multi-step research across sources.Serving and feedback. Answers with citations delivered inside Slack, Teams, the CRM, or the service desk, with ratings and escalations feeding evaluation.Each layer fails in its own way, which is why architecture reviews beat feature checklists. A brilliant model on top of a stale index produces confident answers to last quarter’s questions.
The Retrieval Layer Does the Heavy Lifting Answer quality is decided before the model writes a word. AWS describes retrieval augmented generation as optimizing model output by referencing an authoritative knowledge base outside its training data, and the retrieval layer is where that authority comes from.
Pure vector search stumbles on part numbers, dates, and exact policy names, so production systems combine it with keyword filters and structured metadata. Microsoft’s guidance for RAG on Azure AI Search documents the same pattern, indexed enterprise content queried per request and passed to the model as grounding. Our comparisons of RAG tools and Databricks Vector Search cover the main engine options, and the advanced RAG guide explains reranking and query rewriting.
Where Knowledge Graphs Add Relationships Vectors capture similarity while graphs capture relationships. A knowledge graph links entities such as customers, products, contracts, and owners, so the system can answer questions retrieval alone fumbles, like which client agreements renew next quarter and who owns each renewal.
Graphs also anchor content that never was plain text. When institutional knowledge lives in diagrams, scanned forms, and slide decks, multimodal RAG pairs visual understanding with graph context so a question about a wiring diagram resolves as reliably as one about a policy PDF.
From Answers to Actions with Agents The newest layer turns retrieval into work. Agentic RAG systems plan their own retrieval steps, consult several sources, and check their answers before responding. And where agents differ from chatbots is execution, an agent can file the access request it just explained.
Standards are forming quickly here. Model Context Protocol gives agents a uniform way to reach knowledge stores, a shift we examine in MCP vs RAG , while agent orchestration coordinates multiple specialists over the same governed knowledge layer.
AI Knowledge Management Use Cases Across Business Functions The strongest business cases share one trait, a team that answers the same questions repeatedly under time pressure. Five functions fit that description in almost every enterprise.
Customer Support and Service Support runs on knowledge retrieval at speed. Agent-assist systems surface grounded response suggestions during live conversations, while self-service assistants deflect routine tickets entirely, a pattern we detail in our guide to AI agents for customer support . First-contact resolution rises because the answer arrives with the ticket.
The economics are unusually easy to defend. Every deflected ticket has a known cost, every shortened handle time is measured already, and the knowledge sources, past tickets and product docs, exist before the project starts.
Engineering and IT Runbooks, incident postmortems, and architecture decisions scatter across repositories and chat history. A knowledge layer that retrieves the relevant postmortem during an incident shortens recovery, and new engineers ramp faster when the reasoning behind past decisions is one question away.
The same layer quietly fixes the internal service desk. Password resets, access requests, and environment questions follow documented paths, which makes them prime candidates for grounded self-service before a human ever gets paged.
Sales and Presales Sales teams lose hours reconstructing security questionnaire answers, pricing precedents, and competitive positioning that colleagues already wrote. Retrieval over past proposals and RFP responses turns every closed deal into raw material for the next one.
Presales engineers feel it most during RFP season. A grounded assistant that drafts questionnaire responses from approved past answers, with citations a reviewer can check, compresses a two-week response cycle into days without loosening accuracy controls.
HR and Employee Onboarding Policy and benefits questions arrive in high volume and low variety, which makes them ideal for grounded self-service. Onboarding compresses when role-specific knowledge paths replace the traditional folder of forty documents, and the same foundation supports broader AI workflow automation across HR operations.
Compliance and Legal Legal teams need exact clauses, not summaries of them. Intelligent document processing extracts structured obligations from contracts, retrieval maps them to regulations, and tools like an AI legal document summarizer compress review cycles while keeping citations intact for audit.
Build vs Buy for an AI Knowledge Platform The real choice has three options rather than two. Teams can assemble the stack from open components, buy a packaged platform, or run a hybrid that buys infrastructure while keeping data pipelines, permissions, and evaluation in house.
Kanerika Service
RAG Development Services
Kanerika designs and builds retrieval augmented generation systems over enterprise document estates, with permission-aware indexing, evaluation sets, and governed rollout.
Explore RAG Development Factor Build Buy Hybrid Time to first value Quarters Weeks One quarter, typically Retrieval quality control Full control of every layer Limited to vendor settings Full control where it matters Data boundary Everything stays in your perimeter Depends on vendor architecture Sensitive stores stay internal Ongoing cost shape Engineering headcount Per-seat licenses that grow with users Infrastructure plus targeted licenses Best fit Deep IP and strict data boundaries Standard use cases, small platform team Most mid-size and large enterprises
Table 2: Build vs buy vs hybrid for an AI knowledge management platform.
Four questions settle most debates faster than a feature comparison ever will.
Does the corpus include regulated or client-confidential data? If yes, in-perimeter retrieval or private LLM deployment pushes you toward build or hybrid.Do you need model flexibility? Compare open source LLMs against the commercial leading LLMs before accepting whatever model a platform bundles.Is freshness the point? Retrieval beats retraining for fast-changing knowledge, a trade-off unpacked in RAG vs fine-tuning .Can a smaller model serve most queries? Routing routine questions to compact models cuts serving costs sharply, as our SLMs vs LLMs analysis shows.Whichever path wins, keep the evaluation set and the permission model as your own assets. Vendors change, and those two artifacts are what make switching survivable.
A 90-Day Implementation Roadmap That Holds Up Programs that start with a six-month platform selection usually stall before anyone sees an answer. The pattern that works starts embarrassingly small, one domain, real users, and hard measurement from the first week.
Days 1 to 15. Scope and baseline. Pick one high-friction domain, audit its sources for staleness and duplication, and collect 30 to 50 real questions with owner-approved reference answers. Record how long each takes to answer today.Days 16 to 30. Ingestion and governance. Stand up pipelines with permission mapping intact, and agree retention and ownership rules with the team that runs enterprise data governance .Days 31 to 55. Pilot with real users. Ship a retrieval-grounded assistant to 30 to 50 users in the chosen domain, log every query, and hold a weekly review of failed answers.Days 56 to 75. Evaluate and harden. Score groundedness and citation accuracy against the reference set, close the gaps found in LLM security review, and tune chunking and context assembly using context engineering discipline.Days 76 to 90. Scale the pattern. Add the next domain, publish adoption and deflection numbers, and template the ingestion, evaluation, and rollout playbook so expansion stops requiring heroics.The sequencing matters more than the calendar. Permissions wired in at step two cost days, while permissions retrofitted after a leak costs the program its credibility.
How to Measure Whether the Program Works Knowledge programs die quietly when nobody owns a number. Five metrics keep the investment honest, and all five can be baselined before the first pilot user logs in.
Baselining is the step teams skip and regret. Without a before measurement on the reference question set, the program’s impact becomes a matter of opinion at budget time, and opinions lose to line items.
Time to answer. Minutes from question to verified answer, measured on the reference question set before and after rollout.Self-service resolution. The share of support and internal tickets resolved without human escalation, tracked per domain.Groundedness. The share of answers fully supported by their citations, scored on a sampled basis every release.Weekly active usage. Employees asking questions per week, the earliest warning sign when trust slips.Onboarding ramp. Weeks until a new hire reaches productivity benchmarks, the metric executives remember.The return calculation stays simple. Hours reclaimed multiplied by loaded cost, plus deflected tickets multiplied by cost per ticket, weighed against platform and enablement spend. Programs that clear the bar usually do so on support deflection and onboarding alone, and everything else lands as upside.
Talk to Kanerika
Ready to Baseline Your Knowledge Gap?
Kanerika scopes one high-friction domain, builds the reference question set, and shows you what governed retrieval would change, in a short working session.
Schedule a Demo → Measuring Retrieval Accuracy Before You Scale Most AI knowledge management programs skip a critical checkpoint: proving the retrieval layer actually finds the right answer before rolling it out to thousands of employees. A chatbot that answers confidently but incorrectly does more damage than no chatbot at all, because employees stop verifying answers once they trust the system.
Three metrics to validate before scaling past a pilot team:
Retrieval precision. Of the documents the system retrieves as context for an answer, what percentage are actually relevant to the query? Below 70% precision, the language model is working with noisy context and answer quality degrades even if the model itself is strong.Answer groundedness. Does the generated answer actually derive from the retrieved documents, or does the model fill gaps with plausible-sounding invention? This requires a human-in-the-loop review sample — typically 100-200 queries scored by subject matter experts — before trusting the system on high-stakes questions.Freshness lag. How long between a source document changing and the knowledge system reflecting that change? For policies, pricing, or compliance content, a stale answer is a liability. Real-time or near-real-time indexing matters more than raw model quality for these domains.Integration with existing tools is the second underestimated factor. AI knowledge management rarely replaces Confluence, SharePoint, or a ticketing system — it sits on top of them. The programs that succeed treat the existing tools as the system of record and build retrieval as a layer that respects existing permissions (so the AI never surfaces a document a user could not otherwise access) and existing structure (so search results point back to the canonical source rather than creating a parallel, harder-to-maintain copy).
Kanerika’s implementation pattern connects the knowledge layer to identity and access management first, before optimizing retrieval quality — a technically impressive RAG pipeline that leaks confidential documents to the wrong audience is a governance failure, not a knowledge management win.
Common Pitfalls That Stall AI Knowledge Programs Most failures repeat a short list of mistakes, and every one of them is avoidable at design time.
Watch on YouTube
How DokGPT Transforms Software Learning
A short walkthrough of DokGPT answering questions instantly from enterprise documents and training videos, showing how retrieval-grounded search cuts the time employees spend hunting for how-to knowledge.
Indexing everything on day one. Stale and duplicate content poisons answer quality. Curate the first corpus, then expand with freshness rules in place.Retrofitting permissions. If retrieval can see what the asker cannot, the first leaked salary document ends the program. Access control belongs in the index, not the interface.No named owner. A knowledge platform without an accountable owner and budget line becomes next year’s dead wiki with better search.Trust collapse after one bad answer. Show citations on every response, publish groundedness scores, and give users a one-click correction path.Treating rollout as an IT project. Adoption is a change program. The teams that win embed the assistant where work happens and celebrate answered questions in public channels.Capturing Tacit Knowledge Before It Walks Out the Door The hardest knowledge to manage is the kind that was never written down. Tribal knowledge lives in the heads of tenured employees — the reason a workaround exists, which stakeholder actually makes a decision versus who is copied on the email, the failure mode nobody documented because everyone who joined after the incident just learned to avoid it by word of mouth. AI knowledge management programs that only index existing documents miss this entirely, because there is no document to index.
Three practical approaches for capturing tacit knowledge before an expert retires, transfers, or leaves:
Structured exit interviews mined by AI. Record and transcribe in-depth conversations with departing subject matter experts, focused specifically on undocumented decisions and workarounds rather than generic exit-interview questions. Run the transcripts through the same knowledge extraction pipeline used for formal documents, tagged as a distinct source type so reviewers know it needs verification before being treated as authoritative.Passive capture from existing work. Mine Slack threads, ticket resolution comments, and meeting transcripts (where available and permitted) for the moments an expert explains a nonobvious decision. This is lower effort than dedicated interviews but produces noisier signal that needs more aggressive filtering before it enters the trusted knowledge base.Shadow documentation during the handoff period. When a departing expert has a notice period, pair them with an AI-assisted documentation workflow where the system prompts them with targeted questions based on gaps it detects in existing documentation for their domain, rather than asking them to write a generic knowledge transfer document from a blank page.The retention curve on tacit knowledge is steep: most of what a departing expert could contribute is captured in the first structured session and drops off sharply after that, so prioritize this early in a notice period rather than treating it as a final-week task.
How Kanerika Builds AI Knowledge Management That Teams Trust Kanerika, a Microsoft Solutions Partner for Data and AI with ISO 27001 certification, has been building retrieval-grounded knowledge systems since before they had a category name. The delivery pattern runs in five stages, assess the knowledge estate and rank domains by friction, design the retrieval architecture around existing permissions, build ingestion and evaluation pipelines, pilot with a scored reference set, and embed the assistant into the tools each team already uses.
The product behind much of that work is DokGPT, Kanerika’s document intelligence agent. DokGPT answers questions over enterprise document sets through retrieval augmented generation, with role-based access enforced on every query rather than bolted on at the interface.
The results are documented rather than promised. An investment bank running DokGPT measured 43 percent faster information retrieval, a 35 percent reduction in manual review hours, and 100 percent role-based compliance across its document estate. A member services organization saw similar outcomes when Kanerika deployed a support agent grounded in its policy knowledge, resolving member queries instantly with full context.
Case Study
Instant, Contextual Query Resolution with an AI Member Support Agent
Kanerika deployed an AI support agent grounded in a member organization’s policy knowledge, resolving member queries instantly with full context instead of long email queues.
Read the Case Study → Beyond the product, the delivery details decide outcomes. Kanerika’s teams chunk contracts differently from runbooks because clause boundaries and step boundaries behave differently in retrieval. They score every pilot against a reference question set agreed with business owners before the first user logs in. And they treat the handover as part of the build, training an internal owner on the evaluation dashboard so quality stays measured after the engagement ends.
Practitioner guidance from those deployments is consistent. Map permissions before you embed a single document, because re-indexing is cheap and rebuilding trust is not. Keep humans on the review loop for high-stakes answers in legal and finance. And resist the urge to launch broad, one domain answered brilliantly beats ten domains answered adequately.
Governance, Security, and Permissions for the AI Knowledge Layer An AI knowledge system that surfaces the right answer to the wrong audience is a security incident, not a knowledge management win. Governance for this layer needs its own explicit design rather than inheriting whatever access controls happened to exist on the source documents.
Three governance decisions that determine whether an AI knowledge system is safe to scale:
Permission inheritance, not permission invention. The retrieval layer must respect the access controls already in place on the source system (SharePoint permissions, Confluence space restrictions, CRM record-level security) rather than creating a parallel, separately-managed permission model that inevitably drifts out of sync with the source of truth. If a user could not see a document in its native system, the AI should never surface it, quote it, or answer a question using it as context.Audit trail for what the AI accessed and surfaced. For regulated industries, log which source documents fed into any given AI-generated answer, not just the final output. This is the difference between “the AI said X” and being able to show a regulator or auditor exactly which documents justified that answer.Data classification enforcement at ingestion. Before a document enters the knowledge index, check its classification (public, internal, confidential, restricted) and apply retrieval rules accordingly — restricted documents should require an explicit access grant to surface at all, not just a permission check after the fact.Build this governance layer before scaling past a pilot team. Retrofitting access controls onto a knowledge system already in broad use is far riskier than designing them in from the first rollout, because by the time a permission gap is discovered, it has likely already been exploited or at minimum exposed information to the wrong audience.
Making Enterprise Knowledge Compound Knowledge that cannot be retrieved might as well not exist. AI knowledge management changes the economics of institutional memory, turning documents, tickets, and conversations into answers that arrive in seconds with sources attached.
The winning pattern is now well established. Ground every answer in permission-aware retrieval, start with one domain where friction is obvious, measure deflection and time to answer, and scale the playbook rather than the ambition. Enterprises that treat knowledge as infrastructure will compound it, and the ones that treat it as documentation will keep paying to rediscover what they already knew.
Frequently Asked Questions What is AI knowledge management? AI knowledge management is the use of artificial intelligence technologies such as natural language processing, large language models, and machine learning to capture, organize, retrieve, and apply an organization’s collective knowledge. Instead of returning documents to read, the system understands questions in plain language and delivers direct answers assembled from sources it has permission to access.
How does AI knowledge management work? Content from wikis, tickets, contracts, and transcripts is cleaned, chunked, and embedded into a searchable index with permissions attached. When someone asks a question, hybrid retrieval finds the most relevant passages, and a large language model composes a grounded answer with citations. Feedback on every answer then improves retrieval quality over time.
How is AI knowledge management different from a traditional knowledge base? A traditional knowledge base stores documents and relies on keyword search and manual curation, so it decays as content grows stale. An AI knowledge system reads questions for intent, retrieves passages across many sources at once, synthesizes cited answers, enforces permissions on every query, and improves with usage instead of decaying after launch.
What role does RAG play in AI knowledge management? Retrieval augmented generation grounds a language model in your organization’s own content. Before answering, the system retrieves relevant passages from the indexed knowledge base and passes them to the model as context. This keeps answers current without retraining, reduces hallucination risk, and attaches citations so readers can verify every claim against its source.
Is AI knowledge management secure enough for confidential documents? Yes, when permissions are enforced in the retrieval layer rather than the interface. Production systems map document access controls into the index, so a query only searches content the asker is entitled to see. Private LLM deployment and in-perimeter retrieval keep regulated content inside the corporate boundary, and access should be verified on every query.
Should we build or buy an AI knowledge management system? Most enterprises land on a hybrid. Packaged platforms deliver value in weeks but limit retrieval control, while full builds take quarters. A hybrid buys infrastructure such as vector search and models, and keeps data pipelines, permissions, and evaluation sets in house. Regulated data and deep intellectual property push the balance toward building more yourself.
How do you measure the ROI of AI knowledge management? Baseline a reference set of real questions before rollout, then track time to answer, self-service resolution, groundedness, weekly active usage, and onboarding ramp. Value equals hours reclaimed multiplied by loaded cost, plus deflected tickets multiplied by cost per ticket, weighed against platform and enablement spend. Support deflection alone often clears the bar.
Will AI replace knowledge managers? No, the role shifts from librarian to editor. AI absorbs capture, tagging, summarization, and freshness detection, while people decide what is true, own the evaluation sets, and govern access. Organizations still need accountable owners for knowledge quality, and programs without a named owner tend to fail regardless of how good the technology is.
