Stop Losing the Decisions You Already Made

20 minute read

Pick a decision your business made six months ago. A real one, with consequences: the pricing change, the hire, the client scope agreement. Now go and find it.

Not the outcome. The outcome is everywhere, baked into the invoices and the rota. Find the decision itself: who made it, what the options were, what the losing option was, and why it lost.

In most firms I work with, that search ends in an archaeology dig. Someone scrolls a chat thread from March. Someone else remembers a call that was never minuted. There is an email thread with the client in which, depending on who reads it, two different things were agreed. Forty minutes later the team has reconstructed a plausible version of its own past, and the meeting that needed the answer has moved on without it.

The cost is not the forty minutes. The cost is that a decision you cannot find is a decision you have not finished making. It will be made again, by different people, with less context, and not always the same way. Teams call this revisiting; mostly it is relitigating, and the tell is that nobody in the room can say what the original reasoning was, only that “we discussed this.”

And the same amnesia faces outward. You walk into a renewal conversation holding half a picture of the client: the commitments live in one person’s sent folder, the concern the sponsor raised in March lives in a transcript nobody re-read, and the client, who was present for all of it, remembers everything. A business that cannot retrieve its own past negotiates against people who can. In a small firm this is sharper still, because the memory is usually one person, and that person is in another meeting, on holiday, or gone.

Here is the strange part. Your business already writes everything down. It just cannot read what it wrote. The meeting was transcribed. The trade-offs were argued in a channel. The client confirmed scope in an email. The record of what happened and why exists, nearly complete, scattered across systems that can only search it, not understand it. The problem was never capture. It is that nobody, and until recently no thing, could maintain a readable memory out of the pile.

We know what a cure looks like, because software found one

Engineering has seen this failure before, at scale, and fixed it, and the shape of the fix is instructive even if you never touch a server.

For years the ordinary state of a computer server was hand-configured: patched and tweaked through years of individual changes until, as Martin Fowler put it when he named the pattern the snowflake server, each one was unique, impossible to reproduce, and frightening to touch, because nobody could say which of its quirks were load-bearing. The cure, configuration as code, was not automation. It was memory: the machine’s desired state written down explicitly, versioned, with every change carrying an author and a reason, so that anything could be rebuilt from the record and any drift from the record could be seen.

Your business is in the hand-configured state. Its actual operating rules, the decisions currently in force, are written down nowhere; they are the accumulated residue of meetings and threads, held in the heads of whoever was present, some of whom have left. People are afraid to change things because nobody remembers why they are the way they are. New joiners spend months reverse-engineering context that was said out loud, once.

The cure has the same shape: the decisions in force need a home that is explicit, kept current, and linked to the evidence. Not minutes, which record that a meeting happened. A memory of what was decided, why, and what the deciders rejected.

Writing things down was never the hard part. Keeping them written was.

Let me concede the obvious before someone raises it: keeping records is not new. It may be the oldest technology your business runs. Ledgers, minute books, contracts, the CRM: your firm files the what meticulously, because invoices and contracts carry legal force, and centuries of commercial practice sit behind the discipline. Two defects have survived all of it. The why almost never gets recorded: the reasoning behind the numbers evaporates in the meeting that produced them. And the records you do keep are write-only memory: filed so they could, in principle, be produced for a dispute or an audit, and consulted almost never in the flow of work. What is new is not the record. It is the management of it: a memory that can be interrogated while you work, validated against its own evidence, and brought, whole, to every new decision.

Businesses have tried to close the why-gap before, and here is the strongest objection to everything in this article, made properly: decision logs have been tried, and they mostly don’t work.

Software engineers formalised the idea as architecture decision records back in 2011; Michael Nygard’s motivation was exactly the failure above, that “one of the hardest things to track during the life of a project is the motivation behind certain decisions.” The format he proposed is complete and needs no improvement. And in most organisations I have watched adopt it, the records are written enthusiastically for six weeks and then die. So do decision registers, meeting minutes, and every shared document titled “Key Decisions” that you have ever scrolled past, last edited fourteen months ago.

The failure is not one of virtue. It is one of economics, and the economics run on a rule every household already uses: behaviour follows friction. We lock doors not because locks are unpickable, but because making a thing slightly harder makes it happen much less often. The reverse holds just as firmly, and it is the half that matters here: make a thing slightly easier and people do far more of it. Writing the record is a tax paid at the exact moment the decision feels most obvious and least worth recording: everyone in the room already knows what was agreed, the next job is calling, and the future reader who will desperately need the rationale is an abstraction. Maintenance is worse: decisions get superseded in a corridor conversation, and no process on earth reliably routes the corridor back to the register. A log that is 80% accurate is not 80% useful. The first time it confidently asserts something the team knows was reversed, it is dead, because now every entry needs verifying against the very scrollback it existed to replace.

Any fix that begins “the team just needs the discipline to…” has failed before the sentence ends. The maintenance cost has to fall to nearly zero. That is what has changed.

The memory can now mostly maintain itself

Reading streams of conversation, summarising them against a structure, and detecting that Tuesday’s thread contradicts April’s record is exactly the work that AI language models are now good at, and exactly the work no human has ever sustained. So the shape that works is not “the team writes records.” And, tempting as it sounds, it is not “the machine drafts and a person approves every entry” either. An approval queue is the old failure in new clothes: a memory that grows only as fast as a human can review it is rate-limited by exactly the discipline that killed every register before it, and the queue you skip on a busy Friday is the register dying again.

The shape that works is this. The machine files everything as it lands: each email, transcript, and chat is categorised, filed verbatim, summarised, and folded into the whole, and every record it writes links to the conversations it came from as evidence. The one thing it never does is guess: what it cannot place with confidence it parks, visibly, for a person, because one confidently misfiled client email costs more trust than fifty parked ones cost time. Then the human work starts, and it is moderation, not approval. Every time you interrogate the memory you are also inspecting it: a wrong summary sits one link from its verbatim source, so it is caught in seconds at exactly the moment someone actually relies on that record, and fixed on the spot. The machine re-checks its own work too, on a cycle, hunting contradictions between records and flagging what a newer conversation appears to supersede. Repair runs from both sides.

And the fixing happens at three depths, which is where the judgment lives now. You fix the entry, and that record is right. You fix the rule that produced the entry, the routing that misfiled it or the pattern that let noise through, and that mistake stops recurring anywhere. And occasionally you change your mind about what the memory is for at all: what gets filed, where the walls sit, what should never be automated. The machine only ever works at the first depth, inside rules written at the second, under a frame that stays entirely human. Judgment has not left the system. It has moved up, to where it compounds instead of queueing. When someone finds a wrong record they fix it, because they need it right now. Nobody fixes the rule that caused it, because that helps everyone later and no one today. In a team, put a name on rule-fixing, or it will not happen.

Earlier I said a log that is 80% accurate is dead, and I meant it, so I owe you the reconciliation. That was true of the hand-written register, and the killer was never the errors; it was that nothing distinguished the right entries from the wrong ones, and checking any of them meant redoing the archaeology. Evidence links change the failure mode. A memory whose every entry is one click from the verbatim source degrades gracefully: errors are findable, checkable, and cheap to fix, and each fix tunes the rules. Trust comes from checkability, not infallibility. An automatically maintained memory will contain mistakes; so did the old way, and the old way was almost nothing, remembered by whoever happened to be in the room. Some of the context, checkable and correctable, beats none of the context, which is what you have now. The standard a business memory has to meet is not perfection. It is better than the amnesia it replaces, by a margin you can check.

One class of record keeps a stricter standard, deliberately. A decision record is born proposed and shows that status on its face until a person marks it accepted: the mass of the memory maintains itself, and the entries that bear weight are exactly the ones carrying a human’s confirmation.

I should be plain about the basis for my confidence, because it is testimony, not a study. I run this arrangement in my own consulting business, a firm of the size this article is written for: an AI reads my email, meeting transcripts, and Teams chats, drafts the records, and maintains the memory while I correct it. One live client engagement is currently carrying over fifty confirmed decision records, each with its rationale, the alternatives we rejected, and a link to the conversation that produced it. When a client and I recently differed on what had been agreed, the answer was a link to a transcript, not a negotiation between memories. It is one firm and it is early: if a year from now my own log has decayed into the same unread register I am describing, the economics claim was wrong, and I will say so. I have written up how the system is engineered, and what it took.

What makes a record worth keeping, whoever drafts it: the outcome, in enough detail to act on; the rationale, including the options that lost and why, because the losing options are the part no one can reconstruct later; the evidence link, because a log with evidence is auditable and a log without it is just another document to argue about; and a lifecycle, superseded rather than deleted, because the history of your reversals is worth as much as the decisions.

A business that remembers: the organisational brain

You may have met the personal version of this idea under the name “second brain”: a system for filing your own notes so your head doesn’t have to hold them. What I am describing is that idea at the scale of a business, and the scale changes what it is. Call it an organisational brain: a maintained, evidence-linked memory of what the business said, did, decided, and owes, that its people confirm and everything else draws on. The decision log is the spine of it, because decisions are where memory loss hurts most visibly. But once the machinery exists, the same memory is underneath everything the business currently does from recollection.

You stop working with half a picture of your customer. Preparation for a client conversation changes in kind: instead of an inbox search and a corridor poll, you ask the record. Every commitment either side has made, every concern raised in six months of transcripts, every scope change and the reasoning behind it, current as of yesterday. I walk into calls having asked one question of the memory: what do we owe them, what do they owe us, and what has changed. Client relationships rarely die on a bad decision. They die on a forgotten one: the commitment nobody wrote down, the concern that was raised, acknowledged, and lost. Understanding the customer is not a talent. It is a records problem wearing a relationship costume.

Your evidence becomes content. Every firm says it struggles to produce credible case studies and expertise-led marketing. It is not short of stories; it is short of records of its stories. Once delivery is filed as evidence, content stops being written from recollection and starts being distilled from the record, with every claim traceable to its source. In my own corpus, a single engagement’s filed evidence has so far yielded a case study and four publishable articles, each carrying its provenance. The piece you are reading came out of the same machinery. The marketing asset was always in there; what was missing was the memory to retrieve it from.

Your outcomes become inspectable. A decision record that carries its own signals, the conditions that would tell you it is still right, turns hindsight into an instrument. You can ask which decisions aged badly and whether they share a shape; which client concerns, raised in passing, preceded trouble; where delivery drifted from what was agreed, and when. Most businesses cannot ask these questions at any price, not because analysis is hard but because the reasoning was never captured as data. I am early here myself, and I will not claim results I do not have. What I will claim is the precondition: you cannot inspect what you did not record, and every quarter without a record is a quarter of your own history you have chosen not to learn from.

The hours come back, and they come back to value. The archaeology, the “can you resend that”, the re-briefing, the months a new joiner spends excavating context: none of it is delivery. All of it is the overhead of not remembering. When retrieval becomes cheap, that time goes back to the only thing a client pays for, which is understanding their problem and delivering against it. This is the unglamorous arithmetic of the whole argument, and it is usually a good sign when the practical case is unglamorous.

And notice what kind of consistency all of this buys, because it is not the kind that makes a business rigid. A memory does not make your decisions come out the same every time: circumstances differ, and the judgment stays human. What it makes constant is the context. Nobody decides in ignorance of what was decided before, why, and on what evidence. Not constancy of outcome. Constancy of context. The same firm, carrying the same memory, shows up to every decision, and in work as messy as running a business, that is what consistency can honestly mean.

Which is the part I would put in front of any leadership team deciding what AI is for. The default pattern points AI at tasks: draft the email, summarise the meeting, write the proposal. Fine, and your competitors have the same tools, priced the same, doing the same things. Tasks are one-shot: the output is consumed and the value stops there. Memory compounds. Every conversation filed makes the next answer better, every confirmed decision makes the next analysis richer, and the asset being built, a structured, evidence-linked history of your own business, is the one thing on the market nobody can sell to your competitors, because it is made of your history, not theirs. Point AI at tasks and you rent efficiency. Point it at memory and you build an asset.

The archive is not the memory

Now the warning, because there is a cheap, wrong version of everything above, and it is the version currently being sold.

The wrong version says: connect an AI to your email and your files, and it will know your business. It will not. It will have your archive, and an archive is not a memory. A model pointed at an undifferentiated pile of everything produces confident answers assembled from the wrong context: the superseded decision, the draft that never went out, the concern that was withdrawn. The failures do not look like failures, which is what makes them expensive.

A working organisational brain is engineered, and the discipline has a name: context engineering, deciding what information reaches the model, and the person, at which moment. In practice that means the memory is layered: raw sources filed verbatim as the ground truth, distilled records extracted from them and linked back, and every question answered from the distilled layer with the evidence one link away. The model never swims in the archive; it gets the right slice, scoped to the question, and so does the human reading its output. This, not the AI itself, is where these projects succeed or die, and it is the working half of the argument we make in our AI Adoption Guide: start from a problem you have validated, and treat context as an operating asset that is built and maintained, not a pile the tool will sort out.

Two honesty notes belong here, because a memory this useful invites overclaiming. The record is evidence of your business, not the business itself: the corridor conversation that never touched a filed channel is not in it, and a brain that claims to be complete is lying to you in a new way. And a memory people know is canonical becomes a stage: there will be a temptation to say things for the log and keep real decisions off-channel. The defences are structural, and you can check them from outside: the brain files decisions, never behaviour; it quotes its evidence rather than scoring its people; and it is readable by everyone whose work is in it. In a twenty-person firm nobody files a grievance about a surveillance programme; they just stop trusting you, and then they leave. A memory that is transparent, evidence-linked, and about the work is the version people defend rather than evade, because it is the version that protects them: the reasoning of whoever argued well is preserved, and reversing their decision now requires engaging with it.

Don’t start a habit. Start a small machine.

Here is where this kind of article tells you to start small: hand-write one decision record after your next big meeting, build the habit, grow into a system. I am not going to, for two reasons. The first is that this whole article has argued that the habit is the failure mode, and advising you to start one would be recommending the thing I just spent three thousand words diagnosing. The second is personal. I did not build my system out of discipline. I built it because I do not have the discipline, or honestly the attention span, to file anything by hand for more than a fortnight, and I eventually stopped pretending that a more organised version of me was on his way.

Never write a record by hand. Not even once, not even to try it, because the trying is the trap: it feels virtuous for exactly as long as every dead register ever felt virtuous. The demonstration you actually need takes five minutes and proves the right point. After your next meeting that matters, take the transcript, or the email thread, and hand it to any AI chat tool with one instruction: draft the decision record from this, with what was decided, why, what lost, who was there, and what would change our minds, and point back at the source. Then correct what it got wrong. In those five minutes you have run the entire loop once: the machine drafted, you judged, and nobody paid the filing tax. That is the whole system, at its smallest, and you have just used it.

Then run the two diagnostics that decide whether to act. Count the decisions your team relitigated last quarter: not revisited with new information, which is healthy, but re-argued from scratch because nobody could produce the original reasoning. Each one is an entry the memory would have held, and collectively they are the business case, priced in your most senior people’s hours. Then look at whatever you are spending on AI and ask what it is pointed at. If every tool is a task tool and none of it is building a record your business will still be consulting in three years, you are renting exactly what everyone else is renting, and building nothing.

If the counts are high, make that loop permanent: build the smallest machine that pays the filing tax for you, not a habit that asks you to pay it. One stream, one scope: meeting transcripts are usually the highest-yield stream a business already produces, and your most important client relationship is the scope where lost context bleeds real money. The chain is the same at every size, from the five-minute demonstration to everything else in this article. A source goes in: an email, a transcript, a note, whatever the work already produces. The model does the placing. The source is filed verbatim, because the filing is the ground truth. The knowledge is distilled from the filing, and points back at it. That is the whole design: source, model, filing, knowledge. You moderate until you are happy with the knowledge coming out. Then you scale: a second stream, a second scope, and not before. The engineering questions, what reads the stream, what drafts the records, where the moderation loop sits, are now ordinary buildable things, and that makes this an engineering problem rather than a discipline problem. Which is what is actually new. For as long as anyone has been writing “let’s document our decisions” into meeting actions, the constraint was never knowing what to write. It was that no human sustains the filing, and now none has to.

A decision you cannot find is a decision you have not finished making. Most businesses are carrying hundreds of them, unfinished, waiting to be made again worse, while their client conversations run on half a picture and their marketing writes from recollection. And the raw material for the fix is already flowing through your email, your meetings, and your chat, written down and unread.

Your business already keeps the records. It is time it had the memory.

If you want to work out what an organisational brain would look like for your business, which streams would feed it, where the moderation loop belongs, and what it would unlock, book a call.

Enjoyed this? One click, no account.

Contents

Contents

Open the table of contents

Sign in to make this yours

One free account across every NKD Agility site gets you:

• Bookmarks — save things to read later.
• History — find that page you forgot to bookmark.
• Votes — tell us what was useful, so there's more of it.

Your activity is yours: it powers your account pages, and we only ever analyse totals per item — never what you individually did.

Sign in Not now

Comments

Share

engineering-leadership.hinshelwood.com

Stop Losing the Decisions You Already Made

Your business already writes everything down, in email, meetings, and chat. It just cannot read what it wrote. Why small and medium businesses lose the decisions they already made, work from half a picture of their own customers, and what changes when an AI maintains an organisational brain, a memory of what happened and why, that people correct in use and draw on.

Copy Link

---

LinkedIn WhatsApp Twitter Facebook Email

Subscribe

More actions

• • Subscribe

• ---

• • 

    engineering-leadership.hinshelwood.com

    Stop Losing the Decisions You Already Made

    Your business already writes everything down, in email, meetings, and chat. It just cannot read what it wrote. Why small and medium businesses lose the decisions they already made, work from half a picture of their own customers, and what changes when an AI maintains an organisational brain, a memory of what happened and why, that people correct in use and draw on.

    Copy Link

    ---

    LinkedIn WhatsApp Twitter Facebook Email