Every team that buys or adopts AI tooling eventually has the same awkward moment. Someone says, “The vendor told us model retraining was included,” and someone else says, “I heard it was a paid add-on.” Both people were on the call. Both are sincere. The problem is not honesty but the way the call was recorded afterwards, if it was recorded at all. This article compares three common methods for capturing what an AI or AIOps vendor said on a call, and argues that the least glamorous one wins.
The comparison is deliberately about the record, not the vendor. A good supplier benefits from a clear record as much as you do, and a poor one is exposed by it. Either way, the question of how you write down an agreement is yours to answer.
How we compare: four plain criteria
To keep this analytical rather than opinionated, each method is scored against the same four questions.
- Fidelity. How closely does the record match what was actually said?
- Shared view. Does the other party see it and have a chance to correct it?
- Findability. Can someone who was not on the call locate it in six months?
- Effort. What does it cost in minutes right after the call?
Method one: rely on memory
Memory is the default, because it needs no effort at all. It performs well on the headline, such as “yes, they support our platform”, and badly on the qualifiers that matter: who pays for integration work, what the response window is for a model that misbehaves, and which data is covered by the arrangement.
Research on recall in everyday life is consistent on one point: people remember the gist, then reconstruct details to fit it. That makes memory worst exactly where AI procurement is most fragile, namely in the boundaries of what is promised. On shared view and findability it scores nothing, since the record lives in a single head that may change jobs.
Verdict: fastest, least reliable, and invisible to everyone else.
Method two: internal meeting notes
The second method is a set of notes, written by whoever happened to be taking them. This is a real improvement on memory. Notes capture more qualifiers, and they can be stored in the team wiki. Decision-focused people often add a short action list as well.
The weakness is subtle. Notes are an internal document, so the vendor never sees them. If your note says “retraining included” and their understanding was “retraining once per quarter”, the gap stays hidden until the first invoice or the first support ticket. Notes also reflect the interests of their author, who may have heard the parts they cared about most. A reader who treats notes as proof will occasionally discover that the other side does not recognise them.
Verdict: decent fidelity, good findability, but still a one-sided account.
Method three: a two-line confirmation message
The third method is to send the vendor a brief message after the call. It states what you understood was agreed, names the date by which each next step will happen, and says what you will both do if something changes. For example: “To confirm our call, onboarding covers the monitoring integration and two training sessions. Anything beyond that will be quoted separately before work starts.”
The message fixes the weakness of notes because the other party reads it. A reply of “yes, correct” turns a conversation into a shared record. A reply that says “not quite” is better still, since you have found a misunderstanding while fixing it costs one email.
Fidelity is high because the sender writes while the call is fresh. Shared view is complete. Findability is good if the thread is filed under the vendor’s name. Effort is a few minutes.
The side-by-side result
- Memory. Fidelity low, shared view none, findability none, effort none.
- Meeting notes. Fidelity medium, shared view none, findability good, effort moderate.
- Confirmation message. Fidelity high, shared view full, findability good, effort low.
The message is not merely a tie-breaker. It is the only method of the three that gives the other party a chance to disagree early. Everything else in the comparison follows from that single property.
What an evidence-led team writes in the message
Because AI and AIOps engagements involve moving parts, a good confirmation names the specifics people tend to leave vague. Think of the message as a short answer to five questions:
- What exactly is in scope, including the models, the data sources and the environments?
- What is explicitly outside it for now?
- Who does the next piece of work, and by what date?
- How will cost be measured: per seat, per request, per environment, or by some other basis?
- If the plan changes, who tells whom, and how soon?
You do not need all five every time. For a quick technical clarification, one line suffices. For anything that touches budget or timeline, use all five. Teams that explore the broader decision side of AI adoption, such as the ideas in our piece on how AI shapes everyday business decisions, will recognise that good decisions are easier to defend when the inputs were recorded cleanly.
Where each method fails first
Memory fails at the first handover. Meeting notes fail at the first disagreement, because the other side points out that they never saw them. A confirmation message fails only when nobody sends it, or when it is so long that nobody reads it. That last risk is easy to manage: keep it short, and put the single most expensive assumption on the first line.
There is also a quiet benefit for newcomers. Someone joining the team months later can read the thread and understand why the arrangement is shaped the way it is. For readers building their way into the field, our overview of the skills a new DevOps engineer should develop first mentions communication as a core one, and this is a concrete example of it.
A note on the same habit outside engineering
The pattern is not specific to AI vendors. Any time you hire an outside professional, the same three methods are on offer and the same one wins. If your next arrangement involves a chartered accountant, say for a company set-up or a review of your books, an online chartered accountant finder that filters by place, specialisation and remote or in-person meetings can shorten the search, and the confirmation message works just as well with them.
Putting it into practice
Adopting the habit does not require a new tool. Pick one trigger and one template, and make them automatic.
- Trigger. Any call where money, scope or a date was discussed.
- Template. Two lines: what was agreed, then by when and what happens on change.
- Deadline. Send it before the end of the working day.
- Filing. Keep every confirmation in one shared mailbox label, so colleagues can see it.
After a quarter, review how many confirmations were corrected by the other side. That figure, whatever it is for your team, tells you how much silent drift the habit has been catching. Teams are often surprised that corrections are common even with excellent suppliers, which is the clearest argument for continuing.
The takeaway
Compared on fidelity, shared view, findability and effort, the short confirmation message beats both memory and internal notes, and the margin is largest on the one criterion people forget: letting the other side correct you. It is cheap, it is humble in tone, and it turns a verbal agreement into something you can both point at. The next time a vendor call ends with everyone smiling, spend two minutes writing down what was agreed before the smiles fade.