The Future of Knowledge Inside the Enterprise
Contents
Key takeaways
- Connecting institutional knowledge to AI agents could help companies reuse experience across sales, delivery and decisions, improving continuity as people change roles.
- Value should be measured through approved work, errors and total effort; faster drafts or more proposals alone do not establish business gains.
- Permissions, decision authority and uncertainty must follow information into agent actions, or one mistaken interpretation can spread across otherwise competent systems.
- Start with a bounded recurring workflow, a named owner and a baseline, then expand only when the results justify the operating costs.
Consider 100 comparable cases that together need 100 hours of human effort. Suppose connected knowledge cuts the time spent reconstructing context and preparing drafts. Review then takes longer, and total effort still falls to 60 hours. These figures are illustrative assumptions, not measured results, and the section on released capacity sets them out in full.
Enterprise knowledge management could become a more direct source of business value as organisations connect their accumulated experience to AI agents that can act on it. The way to test that claim is to measure total effort per accepted outcome, along with material errors and later rework.
Customer insights could inform more commercial decisions, and lessons from delivery could improve the next proposal. The reasoning behind a critical decision could also stay available long after the people involved have changed roles. Connected knowledge would change how companies develop new capabilities too, because each completed piece of work contributes context that later work can reuse and refine.
Getting there requires enterprises to connect information across existing systems, distinguish authoritative decisions from tentative suggestions, and define where agents can act independently. Leaders also need to decide how released capacity will create value and how employees will develop the judgment needed to direct increasingly automated work.
I think of an enterprise AI brain (for lack of a better term) as a permissioned memory of the organisation, connected to agents that prepare and execute work under human accountability. Its value would be measured in better business outcomes and the organisation’s ability to retain and apply what it learns.
Four ways connected enterprise knowledge could create value
Faster preparation may release capacity, but the larger opportunity could lie in applying existing expertise to more customers, decisions and products.
Support growth by extending the reach of expertise
A sales team preparing a proposal could draw on comparable projects, the commitments those teams made and what delivery later learned. With that experience easier to apply, the company can pursue more suitable opportunities without a proportionate increase in preparation effort. It can also improve qualification and help teams make commitments they can deliver profitably.
Across authorised customer records, recurring problems could suggest an opportunity for a new service or product, so experience that previously informed a single project can become an input to a broader commercial decision.
Relevant measures include win rate, contribution margin and time to market. A rise in the number of proposals does not, by itself, show growth or improved economics.
Improve efficiency across the full workflow
Connected knowledge could reduce the effort required to find context, reconstruct earlier decisions and recreate existing work. Agents can then use that context to prepare the next task, carrying the relevant evidence and constraints with it.
The unit of measurement should be an accepted outcome. A faster draft that requires more correction may increase total effort. Conversely, an organisation can spend more time on deliberate review while reducing the effort required to complete the workflow. The financial benefit depends on how the released capacity is used and what the system costs to operate.
Improve accuracy by making decisions easier to verify
A proposal linked to its sources, assumptions and approved requirements gives reviewers a clearer basis for judgment. An agent could identify a conflict between a customer commitment and a delivery constraint before it becomes an implementation problem. Changes to an underlying decision can trigger a review of the work that depends on it.
The system can also introduce errors or propagate a mistaken interpretation. Teams should therefore judge accuracy by material error rates and later rework, including errors that automated checks accept at first. Traceability lets reviewers check a conclusion, but a traceable conclusion can still be wrong.
Strengthen continuity by preserving institutional knowledge
A company’s experience includes exceptions, rejected alternatives and the reasons a particular approach succeeded or failed. Preserving that context could reduce dependence on individual employees for explanations that the organisation can reasonably record and share.
New colleagues can then understand why a constraint exists as they encounter it in their work, and experts can spend less time repeating explanations and more time addressing unfamiliar problems. Measures such as time to independent work and the frequency of expert interruptions help assess whether knowledge retention is improving continuity.
Some expertise remains tacit and develops through practice. Recorded decisions can support learning without reproducing the judgment of the person who made them, so the aim should be to make that expertise easier to learn from and apply across the organisation.
Connecting organisational memory with execution
In the operating model I propose, existing systems keep their authoritative records: code in Git, documents in SharePoint or Drive, and customer information in the CRM. A context layer connects those records to the decisions and rationale that explain how they should be used.
This extends the coordination argument in Enterprise AI Strategy is Backwards. The opportunity grows when knowledge captured in one workflow becomes usable in the next, subject to the relevant permissions.
Shared context preserves sources, ownership, access rights and decision history. It distinguishes approved decisions from suggestions, retains unresolved conflicts and shows when an earlier conclusion has been superseded. Knowledge owners are responsible for correcting or retiring stale material, including the summaries derived from it.
Agent execution translates that context into scoped work. Agents can prepare decision notes, propose changes to a backlog, draft a customer response or implement an approved task on a branch. Each assignment has defined permissions, an accountable owner and a record of the evidence used. Models can change without the organisation losing the history around the work. The Jev model-routing project explores this separation at a smaller scale by selecting a model according to the task and available evidence.
Human judgment establishes objectives, allocates authority and resolves consequential uncertainty. Review should give people access to the proposed change and its supporting evidence. Disagreements between automated reviewers show people where to look, while inspecting accepted outputs helps find shared mistakes.
In a hypothetical design review, the team agrees to simplify an onboarding flow, postpone a feature and retain a customer exception. An agent prepares a decision note for confirmation. Once the relevant decisions are approved, another agent checks the backlog, proposes ticket updates and identifies an older instruction that conflicts with the new scope.
Work that falls within existing permissions goes ahead. Coding agents prepare changes and run tests; reviewers check them against the approved task. The conflicting instruction returns to the project lead. By the next morning, the team has proposed changes to inspect and a specific question to resolve, with a record linking both to the original decision.
The value comes from continuity across the workflow, because a decision can inform later work without each participant having to reconstruct its meaning. That puts a matching responsibility on the context layer. A suggestion recorded by mistake as an instruction could spread through several otherwise competent agents, so authority and uncertainty must travel with the information.
Converting released capacity into business impact
The value of released time depends on what the organisation decides to do with it.
For the 100 comparable cases in the opening, assume employees currently spend 35 hours reconstructing context, 35 preparing drafts, 20 reviewing them and ten correcting them. Now assume connected knowledge reduces reconstruction to ten hours and preparation to 15, while review rises to 25 and correction remains at ten.
These figures are assumptions, not measured results or a forecast. They hold case volume and required quality constant, and show how a reduction in preparation can make room for more review and still release 40 hours of capacity.
That capacity could support additional customer work, faster delivery or activities that improve future performance. It reduces spending where it replaces overtime or an otherwise necessary external expense. If it is absorbed by additional low-value activity, the financial benefit may be negligible.
An investment case also needs to include technology, implementation and ongoing operating costs, which are excluded from this example. Maintaining context, reviewing exceptions and correcting agent errors all consume resources. Measuring total effort and business outcomes over time shows whether the released capacity is worth more than those costs.
Establishing the conditions for dependable scale
Indexing confidential client work, employee information or commercially sensitive plans must not make them accessible to everyone. Permissions need to follow information into summaries and proposed actions as well as govern access to the original record.
Decision authority also needs to remain explicit. A recent message does not necessarily override an older approved requirement. An agent authorised to prepare a draft may have no authority to send it to a client. A team can expand autonomy for a well-understood class of work as evidence accumulates, while keeping that permission specific and revocable.
Review capacity is another design constraint. Agents that produce proposals continuously could create more work than people can assess. Priorities and limits on work in progress should therefore be part of the operating model. A low-cost proposal can still be expensive to evaluate, and the system should be able to defer work whose value does not justify that attention.
Automated review can support this process, but agreement is not independent verification. Two agents may reach the same conclusion because they relied on the same mistaken decision note. Human reviewers need access to underlying evidence and periodic samples of work that automated checks accepted.
Recovery should be defined according to the action. A document can often be restored from version history; a sent email cannot be recalled with the same certainty. Pausing an agent can stop its next action without reversing the consequences of its last one. Teams need to know which changes they can undo and which require repair.
Employees build judgment partly by preparing analyses, implementing changes and seeing where their reasoning fails. As agents take on more of that preparation, organisations should create deliberate opportunities for people to form their own view, inspect failures and take responsibility. Shared decision histories can help by letting people compare alternatives and their outcomes.
A leadership agenda for enterprise knowledge management
A practical starting point is a recurring workflow with a measurable outcome, a named owner and a bounded body of knowledge. The aim is to find out whether connected knowledge improves the work enough to justify broader adoption.
Decide which outcome the workflow should improve, whether delivery speed, rework, commercial growth or continuity, and establish a baseline that includes preparation, review and correction. Choose work whose result can be evaluated at a consistent quality standard. Identify the sources, decisions and exceptions that inform the work, and assign responsibility for maintaining them. Make access rights, approval requirements and unresolved conflicts explicit before agents act on that context.
Test continuity as well as productivity. Check whether someone unfamiliar with the project can understand the reasoning and carry the work forward, then change an underlying decision and see whether the system identifies the affected tasks. These tests reveal whether the organisation is retaining usable knowledge or accumulating more records. Compare accepted outcomes, cycle time, material errors and total operating effort before widening the scope, and connect successful workflows where permissions allow it, so delivery experience can inform sales and customer feedback can inform product decisions. Expand only as fast as the organisation can maintain the knowledge and judge the resulting work.