7 minute read
- Manual filing could not maintain a usable history of NKD Agility’s consulting engagements alongside ongoing delivery work, leaving key reasoning and commitments effectively unavailable.
- Separating deciding from recording allows Engage to capture agreed decisions automatically while humans retain responsibility for judgement about the decisions themselves. see why — jump to the section that argues this
- Engage’s value depends on what communications Martin can capture and on his ongoing validation and rule-tuning; the system cannot independently guarantee correct or complete records. see why — jump to the section that argues this
- Start from a recent agreement you had to reconstruct and test whether you can find what was decided, why, and the supporting conversation or correspondence.
- Check whether you can follow how that agreement has changed over time, so you see what needs capturing and ongoing maintenance.
- Let an automated system handle routine filing and extraction, but keep responsibility for judging what records mean and whether you can rely on them.
A client relationship should rest on what was actually agreed, not on whoever remembers it best. I run my consulting business on AI agents I do not trust to be right on their own. What I trust instead is the record they keep, because almost every decision in it links back to the email or transcript it came from, and anyone I share it with can check it.
In August 2026, a client manager at Client A asked whether their business was getting the consulting time it had agreed. Our expectations did not match.
“I was horrified that there was a mismatch between what they understood and what I said. It means what I said, and what they signed, was not clear, or they did not read it.”
What had been said was on record. The discovery transcript held my original explanation: about a quarter of my time, averaged across the month. I restated the 25% arrangement, and on the next day’s call the client’s delivery lead acknowledged the mismatch and said my response was appropriate. The record also showed that the proposal never stated the 25%. It was meant to be there but had been lost in an edit, so the client accepted a proposal that described non-exclusive availability within a fixed timebox and nothing more. I have since put it back into my proposals. The record did not settle who was right; it showed what had actually been said, and it caught my mistake as well as theirs.
Why it mattered
Holding to what was agreed is hard across several client programmes. Left alone, the gap shows up as a mismatch over agreed scope with a client who has already accepted the proposal.
“Everything got lost. What they said, who I spoke to, who they were. Coming back around and talking to the same person was often a new conversation, not one that’s built on what went before.”
Consultants and delivery leads working across several programmes face the same problem. For their clients, a shared record means agreements that can be checked against the source and decisions that survive a change of stakeholders.
Why agents
The record has to be kept up to date and it has to be checkable. The alternatives were doing nothing, paying someone to take notes, or keeping a system up to date myself, and people do not keep records up to date.
“I’ve tried to use a CRM many times, and what gets lost is that someone has to keep it up to date. Someone has to take notes, write up those notes, and then read those notes before the next conversation. That’s not what happens.”
Agents are probabilistic, not deterministic, so they are prone to the same class of judgement errors as humans, and that needs guardrails. Unlike people, they will keep the record up to date, so the record has to carry its own evidence. That is why I built Engage, an organisational brain that agents maintain, with its first commit on 28 May 2026. I record and transcribe my meetings so there is something to capture. Engage takes in email, meeting transcripts and the Teams chats I have permission to capture, files each one against the right engagement, and turns them into decisions, actions, meetings, people, insights and risks, most linked to their source. The agents do the filing; I define the rules and change them when I notice they produce the wrong result, which happens only for records I use.
Engage records decisions; it does not make them. When a transcript shows a decision was agreed, Engage records it as accepted without asking me to approve it again, a change from the approach I described in Stop Losing the Decisions You Already Made. Uncertain evidence is flagged for me to assess.
What changed
The source, not memory. The spoken explanation behind Client A’s question was preserved in the recorded transcript, filed against Client A, and the gap in my own proposal is now fixed.
Reasoning carried into the next piece of work. In September, a migration-planning discussion at Client B separated two concerns: the risk that build preparation would take too long, and doubt that the work could be done at all. The decision record kept the resulting ordering and its reasoning, and a shared draft risk register that cited it was included in the stakeholder presentation. Three days later the team returned to build readiness, and I added build-team and dashboard rebuild work to the plan.
Knowing who people are. Before a meeting with someone I have not met, I can see their organisation and role where they are on record, and for 82 of 90 people a job title.
“I know the context. I can pull the context for a client before each meeting. I can analyse what they say to find things that need asking, actions, or inconsistencies we need to discuss.”
How far the record can be trusted
At an inspection on 13 October 2026, Engage held records for seven client engagements: 135 decisions, 249 actions, 105 meetings, 143 conversations, 136 insights, 90 people and 10 risks, drawn from 515 emails, 109 meeting recordings or transcripts and 4 Teams chats.
134 of the 135 decisions linked to a source, and for 133 of them the linked file exists. That proves the file is there; nothing checks automatically that it supports the record. Six replaced decisions point to the decision that replaced them, and five of those replacements point back, so earlier reasoning stays available. The chain runs from transcript to meeting, decision and action; which deliverable an action produced is implied by naming, not recorded.
Some controls are enforced, but only on what the agents write: automatic checks block agents from changing Engage’s own rules, check that records are correctly formed, and check that proposals stay within their agreed timebox. I have seen nothing that restricts which engagements an agent can read. Other controls are standing instructions, such as asking before any change, which guide the agents but prevent nothing. Engage has a stated retention schedule, keeping engagement content for the life of the engagement plus seven years; I have seen nothing that enforces it.
What it costs and still misses
The evidence chain protects against the agents’ mistakes, not against attacks. Checks find malformed records; they cannot tell whether a record’s meaning is right or whether it was filed against the right engagement. I check meaning and filing as I use the records, so a record nobody uses is never checked, and because the checking depends on one person, the least-used engagements are the least checked. There is no set process for repairing a wrong or misfiled record; when I find one, I go back to what was brought in and correct it from there. Every client’s material sits in one record, so a misfiled record is a confidentiality risk as well as an accuracy problem.
Captured email and transcripts are untrusted input, and I have seen no defence against instructions hidden in that content. Because agreed decisions are recorded as accepted automatically, a crafted email could produce a decision that looks agreed. Nothing prevents it; the link only lets someone trace it afterwards, if they check. How participants consent to recording, and how deletion requests fit the retention schedule, are not documented.
Coverage is uneven: one engagement holds about half of all records, team records exist for two of seven engagements, no record captures a stakeholder’s influence, no decisions about running the business itself had been recorded at either inspection, and chats inside customers’ own Teams environments are not captured. Capture completeness has not been audited, and retrieval is not logged.
The costs are model usage, my time checking records, and fixing rules when they go wrong. None of these, nor the time saved or the financial return, has been measured.
A starting point
Take a recent agreement you had to reconstruct. Can you find what was decided, who decided it, why, and the conversation behind it? If a new stakeholder questioned it tomorrow, could you show them the source?
Bring one agreement you had to reconstruct to NKD Agility AI Systems, and we will show you what an evidence-linked record of it looks like, kept by agents whose work you can check, so the agreement rests on what was said, not on who remembers it.
Was this useful?
Enjoyed this? One click, no account.
Questions this answers
How can I keep business decisions, agreements, and their reasoning easily available in my day‑to‑day consulting work?
By capturing emails, meeting recordings, and transcripts, and then using a system like Engage to automatically extract, file, and link decisions, agreements, and actions back to their original sources, you can maintain a usable history of each engagement. This lets you work from concise decision and commitment records while retaining the full transcript or email thread for reference when details are unclear or disputed.
What are the limitations and responsibilities when using an AI system to track and maintain business and client decision records?
The records are only as good as the inputs you can capture and supply, and some channels like customer‑hosted Teams chats may remain gaps unless manually exported. You must still verify meaning, engagement assignment, and correctness, adjust the system’s filing and extraction rules when it misstates agreements, and accept the ongoing cost and time to run and check the system, because it does not remove your responsibility for the work produced from those records.