The Deliverable That Wasn’t on the Statement of Work
How I taught myself to build the talking deck.
Strategy consulting sells insight and then files it where no one will ever find it again.
Every project ends the same way. An eighty-slide deck, a folder of transcripts, two or three competitor teardowns, a market model held together by hope and merged cells. The client calls it invaluable, saves it somewhere sensible, and a month later asks a question that was answered on slide 43. The insight was never the problem. Recall was. The whole thing was built for a one-hour readout, not for the year afterward when someone actually needed it.
So one weekend I tried something that was on no statement of work anywhere.
The first one ran on a password and not much else
I took the research deck from one engagement, a study for a hardware client, and wired it to a language model so you could simply ask it things. What did buyers say about cost? Give me the three adoption barriers. It answered in plain English, from our own research, instead of making anyone scroll.
I put it behind a password so it wasn’t loose on the open internet, then tested it the way a client would. It broke on the first table I asked for. I fixed it. It broke somewhere else. I fixed that too. Most of building with this, I have learned, is precisely that loop.
When it worked, though, it was a different object than a deck. You didn’t read it. You questioned it, and it answered out of work we had already done.
Nobody asked me to build it, which is the whole point. There was no line for “consultant teaches himself retrieval-augmented generation on a Sunday.” I had a hunch that our most valuable asset was trapped in our least usable format, and the only way to test a hunch like that is to build the ugly first version and see.
It stopped being a one-off
Once one existed, everyone wanted their own.
The competitive-intelligence project got one, which felt fitting. A healthcare group-purchasing engagement got a version the client could query directly. Then the printing research, the channel-partner strategy, the large-format work, an OEM-intelligence corpus, the workstation studies. It ended up being roughly eight of these, each pointed at a different mountain of research, and somewhere in there I started buying proper subdomains, because “another-bot.some-host.space” is not a link you send a Fortune 500 client with a straight face.
That was a bigger jump than it sounds. A prototype you demo to your own team can be rough, and everyone forgives it. A tool a client uses for real work cannot be, and closing that gap is most of the job.
The one that convinced people
The version that settled the argument was a competitive-research bot, fed every transcript and deck from the project. One of the client’s own customers, someone we had never spoken to, used it to draft a creative brief. In German, for their agency.
So a stranger, several steps removed from us, used a thing I had built to make their own deliverable, in a language I don’t speak, off the back of our research. That was the moment I stopped calling these chatbots and started thinking of them as a deliverable in their own right. Not a deck with a bot bolted on. The research itself, awake enough to answer, with the deck demoted to a summary.
“The research itself, awake enough to answer, with the deck demoted to a summary.” THE LINE THIS WHOLE SECTION GREW FROM
What I took from it
The technical part mattered least. I picked it up as I went, and so can anyone. The fiddly bits, chunking and retrieval and the elegant ways the thing gets confidently wrong, are in the next piece.
What mattered more was noticing that a deliverable everyone treated as settled had a better version hiding inside it, and being willing to look slightly foolish building the first draft to prove it.
Consulting rewards the person who answers the question they were asked. Every so often it pays to answer the one nobody thought to ask: what if the research could answer back? Worst case, you lose a weekend. This time it added a new line to the menu, and it started with a password and a bot that couldn’t draw a table. ▮