SaaS AI platform
The company that sold what it did not use
A SaaS company of about 120 people was building an AI platform, selling it well, and running internally on project boards, shared docs, and people retyping information between systems. The sales team demonstrated agents doing the work of whole departments every week, then closed their laptops and went back to a company that did not operate that way. The CEO named the gap before we could: figure out AI for our operations, not just our product, before anyone outside the building notices.
What was at stake
Demand was growing faster than leadership wanted to grow headcount, and the internal teams, customer service, the SDRs, operations, and marketing, were carrying enough coordination load that the growth had nowhere to land. The credibility risk was real but secondary. The operating risk was that the company's own scale would stall against its own internal friction, and that the people already on payroll would keep swimming in different directions while leadership debated which roles to add.
What the floor revealed
We proposed the standard entry: a fixed fee, a day inside the operation, a playbook within two weeks. The floor was made of screens, so we sat with the people who lived in them.
A senior support lead answered a configuration question in four minutes by pasting from a personal spreadsheet of answers she had written before, polished over years, organized by topic. It was the third time that week for that question. The spreadsheet functioned as the company's knowledge system, but one person owned it, and the company did not know it existed. A services project manager spent most of her Friday building the weekly status report by hand across her share of the team's [dozen] live implementations, reading boards, chasing threads, reconciling the two, writing the summary the customer expected. "Most of my Friday is producing this," she said, about a report describing work she had not had time to do because she was describing it. A standing sync's entire timed function was reading a board aloud to people who did not trust it was current. Marketing and sales presented two versions of the same pipeline.
Every piece of that load was work the company's own product does for its customers. The platform was sufficient. The platform was, in fact, down the hall. Nobody owned pointing it at the building it was built in. That is an ownership gap, not a technology gap, and it does not show up on a dashboard.
The sharper finding was the loop underneath. We traced one larger customer from signed contract to steady state. The quietest step paid for everything downstream: when customers' users learned the product, they got a recorded webinar and a documentation link, and were structurally on their own. The queue inherited what self-teaching missed. When we pulled the queue's exports apart, [roughly half] of the volume was previously answered questions. Re-answering consumed [most of two] full-time people. The four teams' coordination work added up to [several] full-time equivalents already on payroll. Leadership was reading the queue as weather and the capacity problem as a hiring problem. The floor said both readings were wrong.
The strategic shift
Instead of treating operations as a separate automation problem, the company would become its own customer first. Operations would run on the platform, then customer success. The trade-off was a longer arc with no quick external win, but the internal deployment would produce the operating knowledge the company needed and the artifacts that would carry the next move.
The second shift was the move the trace had earned, the one that broke the loop instead of servicing it. Teach the customers, not with more documentation, but with the operating knowledge the company was about to author on its own floor. The training program would be extracted from the doing.
How we built it
The first build went where the blast radius was private: the services operation. Digest agents, running on the company's own platform, read the boards, the threads, and the calendars together, producing each morning the truth project managers used to assemble by hand on Fridays. The old motion ran beside the new one on purpose until the comparison got boring. Then the operations lead retired the Friday report and the first standing sync, by logged decision, which is how you retire a motion without leaving half of it running out of habit. Board hygiene and handoffs followed. The work management platform and the office suite stayed; agents carried information between them so people stopped living logged into either.
The queue was the harder promotion because the words face customers. Before any of it went live, we ran the cheap test: agents drafted answers to the recognized questions from the documentation and the closed-ticket record, offline, judged blind by the senior support lead against what she would have written. The drafts were good. That was not the finding. The finding was the pattern in the misses, every so often the machine answered confidently, fluently, and correctly for the previous release. That catch, made by the one person whose spreadsheet had been keeping the company honest, became the architecture: the knowledge layer was rebuilt with product-version controls, her library was codified into it, and the rule was written before a single customer saw a machine-drafted word.
Then the human gate went up. Every customer-facing draft passed through approval, every catch was logged with its pattern, every pattern became a rule. The catch rate fell week over week until recognized answers earned lighter review on evidence, the same way the company asks its own customers to graduate their agents.
Governed turned out to be a different state than live, and the difference announced itself quietly. Months in, a team restructured its boards, ordinary housekeeping, and the digest reading them kept producing confidently from the stale map. Nothing errored. The numbers stopped moving, and the operations lead caught it on the weekly read because a project everyone knew was sprinting looked still. Every reading agent got schema checks and a standing rule, halt and ask rather than produce from a stale picture. The scar became how the company now teaches the discipline.
What changed
The company recovered capacity inside the teams already on payroll. It runs on its own product every day, in operations and in customer success, and the sales demo is no longer a demo, it is a tour. The capacity leadership thought it would have to hire was already on payroll, and the coordination work that had been absorbing it has been retired or compressed.
The operating change is the deeper result. The known half of the queue now drafts by machine in minutes. The senior hours that easy tickets used to consume have moved to the escalations and judgment calls where they belong. The support lead went from writing answers, to judging the machine's, to teaching, which is the job her spreadsheet had been trying to be all along. Customer ramps compressed, and the trained customers lean lighter on the queue, which gives the senior people their judgment hours back. The loop the audit named now runs the other way.
The training program is the artifact that closed the arc. The playbooks, gate designs, graduation rules, versioning rule, and the stale-board scar were authored into an operator's education that teaches customers how to run AI inside their own businesses. The curriculum's authority is its provenance: every lesson in it is something the company demonstrably does, because the course was extracted from the doing. It piloted with [a few] customers, was revised on their feedback, and is standard for new ones.
What stands as proof
The Day One Audit recorded the operating baseline before any build and named the loop. The Playbook sized the five moves and sequenced operations first, customer success second, the training program third. Each major decision is logged in the Charter with its alternatives, its rationale, and the evidence behind it, including the offline drafting test, the read-only digest trial, the retired Friday report, and the stale-digest incident that earned the halt-and-ask rule. The training program now teaches customers using the SaaS company's own working records as the curriculum.
For an operator running an internal team that is buried while the product ships fine: more headcount will not move a load made of coordination work the product itself can do. Run the company on its own platform first, gate the customer-facing words with the person whose judgment is already keeping you honest, and let the operating knowledge you author internally become the thing you sell next.