Executive summary & recap:
In Part 1, we explored why enterprise AI rollouts stall at a 14% adoption rate when treated like standard software procurement. Frontier AI value lives in the distance between the model and company data. In Part 2, we map out the four structural components of the Deployment layer: Deploy (governance), Ground (MCP & business rules), Adopt (training on live data), and Build (production agents). We demonstrate through the case of Bora Brasil why grounding must always precede training.
Between the model and the business sits a layer of work that is neither pure IT nor pure change management. I think of it in four parts, and I hold a strong view on each.
1. Deploy
The platform itself — Claude Enterprise, Claude Code and Cowork, or ChatGPT Enterprise and Codex — rolled out properly. In practice that means a licence model that allocates seats by role, area and seniority rather than first-come-first-served; identity through the corporate SSO with group-based access; a usage policy that says in plain language what may and may not be pasted in, and a best-practice guide that says what good looks like; and cost governance that shows, per department, who is using what, from the first week. This is the part most organisations do reasonably well, and it is the least valuable part on its own. A perfectly provisioned tenant with nothing connected to it is a very expensive chat window.
2. Ground
The context layer that makes the model answer with the company’s own truth. Concretely: a single MCP server, running in the client’s own cloud tenant, that exposes the structured data (the sales tables in Fabric or Snowflake, the budgets, the KPI marts) and the unstructured repositories (contracts, policies, board decks and procedures in SharePoint or Drive) through one governed door.
On top of it, a business knowledge layer: the official glossary, the KPI definitions as Finance writes them, the business rules (“a quarter closes on the 25th”, “exclude intercompany from net sales”), and a semantic description of every document so the model can find the right file and read it in full rather than guess from a snippet. Access mirrors each employee’s existing permissions, so the model sees for you only what you could already open yourself. And every answer carries a citation: the table, the row filter, the document, the page.
I believe this is where enterprise AI is won or lost, and it is the part almost everyone skips because it looks like plumbing. Without it the model is a brilliant stranger. With it, it is a colleague who has read everything and can prove it.
3. Adopt
Training, department by department, on live company data. I have become allergic to the generic “prompting masterclass” with its sample emails and its imaginary company: it produces a week of curiosity and then silence. The sessions that work look different.
- Sales spends two and a half hours on one real question about last quarter.
- Finance reconciles a real month-end file.
- HR rebuilds a real absenteeism report.
- Legal reviews a real NDA against the company’s own playbook.
Each participant leaves with a Claude Project that holds their department’s rules and a working artefact they will open again on Monday. The difference is not pedagogy. It is whether the model was grounded before anyone was trained on it.
4. Build
Production agents, developed in short sprints by engineers who sit with the client’s team, on a framework the client keeps. The unglamorous work — reading supplier invoices into a schema and flagging unmatched lines, drafting meeting minutes and actions from a Teams transcript, checking an NDA against the house guidelines and summarising the liability and termination clauses, extracting the numbers from a stack of PDFs into the underwriting template — is where the hours are, and it is where most agent programmes never arrive because they start with something impressive instead of something useful.
These four are a system, not a menu, and I would go further: doing one of them without the others is often worse than doing nothing. Training without grounding teaches people that the tool hallucinates, and they are right. Grounding without training builds a cathedral nobody enters. Agents without a platform produce a dozen disconnected automations with no shared governance, and the security team eventually, rightly, shuts them down.
Example: Who answers the sales question at Bora Brasil?
Consider a simple question a commercial director might ask on a Monday morning: “How did Skala perform year-over-year last quarter, and where did we lose share?”
At Bora Brasil, the haircare group formed from Skala Brasil and Lola From Rio, that question used to travel. From the director to an analyst, from the analyst to a BI extract, from the extract to a spreadsheet, from the spreadsheet to a deck, and back to the director three days later, by which point the follow-up question had changed. Meanwhile fifteen early adopters were already using Claude informally, connected to their email, calendar, Teams and files, getting plausible answers that nobody could trace to a source and that sometimes used a definition of “share” nobody in Finance recognised. Leadership’s three requirements were exactly the three questions above: answers grounded in the group’s own data, access that respects existing permissions, and a governed environment so that sensitive commercial information stopped drifting into unsanctioned tools. That is the 14% moment, seen from inside.
What changed over nine weeks was not the model. Three streams ran in parallel: a bottom-up adoption stream, a top-down roadmap stream and a governance stream. The engineering work produced a context platform on Azure, approved by the group’s own technical leadership, that exposes Microsoft Fabric (sales, KPIs, budgets) and SharePoint (contracts, policies, presentations) through a single MCP server. A business knowledge layer holds the glossary, the KPI definitions and the business rules, with a semantic description of each document so Claude can locate and read the full file. Access mirrors each employee’s existing permissions, and every figure cites its source table or document. In parallel, interviews with the directors of Commercial, Marketing, Innovation and R&D, Operations, Finance and HR produced 83 opportunities, scored on business value and technical complexity, narrowed to 24 and then, in a leadership workshop, to five for the year-one roadmap.
Then — and only then — came the training, on exactly that sales question. In a 150-minute session each participant set up a Claude Project with the business rules, pulled the KPI definitions from SharePoint, planned the query before running it against Fabric, built an interactive sales dashboard as a shareable artefact and checked the source of every number. The faster ones turned the dashboard into a weekly scheduled routine in Cowork before the session ended. The facilitator’s guide, itself built with Claude, tracked each participant’s progress step by step.
Sixty-two employees are live today, scaling to 150, with licences allocated by role and area, an AI Office at the centre and a Champion in each business area. The first thirty licence holders report at least four hours saved per week on average; seven in ten save three hours or more. One HR payroll data-load conversion took 90 minutes instead of the 40 consultant hours quoted externally. A timekeeping report that used to take eight hours of manual checking now runs on a schedule.
I tell this story not because the numbers are large — they are early and self-reported — but because the sequence is the lesson. Ground first. Train on the grounded thing. Hand the habit to someone who stays. Reverse the order and you get a webinar.
⏭️ Up next in Part 3 – The forward deployed operating model
Building and grounding this system cannot be left to traditional steering committees or external slide decks. In Part 3, we detail the forward deployed operating model—how pairing strategists with engineers in the building creates lasting capability, the risks of skipping it, and how Chancerygate built 20+ production agents independently.
