Last time we argued your next CRM shouldn't be a CRM. The data model belongs in a firm-owned warehouse, the workflows belong with agents, and the UI belongs to you. The principle landed. The feedback, almost without exception, was the same: agreed, now show me how. Fair challenge. Here is the how.

A caveat first, because twelve weeks does not retire your CRM. What twelve weeks does is get the headless pattern live in production: the warehouse stood up, the CRM demoted to a feed engine, the first agents doing real work, and a pilot team off the CRM screens for good. The full transition still runs six to nine months, but tangible benefits are delivered within the twelve-week sprint.

What you are actually building

Strip the programme to its essentials and there are four components:

  • A governed data warehouse that becomes the single source of truth.
  • An MCP gateway through which agents read and write to every system, the demoted CRM included.
  • A small portfolio of agents handling the non-client-facing work.
  • A thin interface, portal, app or chat, that replaces navigation with conversation.

None of this is exotic in 2026. The warehouse is config, not code. MCP connectors already exist for the major custodians, platforms and back-office systems. The frontier model vendors have done the hard part. What is left for firms is sequencing, governance and adoption. That is what the twelve weeks buys.

The critical path is data, not technology

Every firm that attempts this learns the same lesson: the agents are the easy bit. The data dependencies are where programmes live or die. So name them on day one, before anyone writes a line of anything.

  • Client master data. The CRM has been your de facto client record for a decade, and it is full of duplicates, stale addresses and free-text notes. The warehouse needs a golden record with survivorship rules: which source wins for each attribute, and who arbitrates.
  • Custodian and platform feeds. Valuations, transactions, fees and cash. Most arrive clean via API or daily file. The awkward residue — legacy products, offshore bonds, paper-based providers — needs a manual workaround that is honest about its cost.
  • Advice and suitability records. Agents drafting client communications must retrieve the latest fact-find, risk profile and suitability letter. If these live in network folders with inconsistent naming, they need indexing into governed, retrievable storage.
  • Communications history. Emails, call notes and meeting transcripts give agents their context. This is the highest-value and most sensitive feed, so permissions must travel with the data.
  • Identity and entitlements. When an agent acts, it acts as someone. Adviser-level permissions are enforced at the MCP gateway, not assumed inside each tool.
  • Reconciliation. The finance team's fee and price reconciliations are the workflows keeping the CRM alive today. They move to the warehouse early, or they anchor you to the old stack indefinitely.

Note what is absent from that list: anything to do with AI. The model does not care whether your data is messy. Your compliance officer does.

Weeks 1-3: foundations and the data audit

Stand up the warehouse and the MCP gateway in skeleton form. Run the data audit against the six dependencies and grade each feed: clean, fixable or manual. Pick the golden-record rules. Choose two launch workflows, not ten; meeting preparation and fee reconciliation are the usual winners, because one buys adviser goodwill and the other buys finance goodwill. And open the pricing conversation with your CRM vendor now. You will be reducing seats whatever happens, the negotiation takes longer than the build, and the arm-wrestle will be fierce.

Weeks 4-6: the plumbing

Land the custodian and platform feeds in the warehouse and reconcile them against the CRM until the numbers agree two weeks running. Connect the CRM itself through the gateway, read-only at first, so agents see everything the old system knows without anyone touching its UI. Index the document store. Wire identity through the gateway so every agent action is attributable to a person. This is the unglamorous phase. It is also the one you must not compress.

Weeks 7-9: agents, with the controls built in

Build the two launch agents against the governed data, with human-in-the-loop approval on anything client-facing. Set per-workflow token budgets and observability from day one — this is how the token bill stays a fraction of the licence bill it replaces, as we discussed last time. Write the evaluation suite before go-live, not after: a library of real cases each agent must pass, rerun on every change. The FCA won't ask if you used AI. They will ask how you evidenced control of it.

Weeks 10-12: the pilot pod and the behavioural reset

Pick one adviser pod — an adviser, a paraplanner, an administrator — and take the CRM screens away. They get the portal, the agents and one of our Forward Deployment AI engineers as a support contact. You measure everything: preparation time per meeting, reconciliation effort, queries the agent answered versus escalated. The pilot exists to manufacture two things: evidence for the board, and advocates for the rollout. A pod that gets hours back each week sells the change better than any steering-committee deck ever will.

We said last time that the biggest barrier is cultural, and the twelve-week shape is built around that truth. People who have logged into a CRM for years will not stop because an architecture diagram told them to. They stop when the alternative is visibly better and saves them work. This is a flightpath rather than a switch.

What week 13 looks like

By week 13 you have a governed warehouse carrying your platform feeds, two agents in production with evidenced controls, and a pilot pod that will not give the new world back. The CRM contract negotiation now happens from a position of strength — enterprise data-and-feeds pricing, or a seat count you can hold up on one hand. From there, the path to full retirement is a planned absorption, workflow by workflow, on your schedule rather than the vendor's.

The firms that struggle will be the ones that ran this as an IT project. The firms that compound the advantage will be the ones that ran it as an operating-model change with a twelve-week ignition sequence. The tsunami we wrote about last time has not slowed; the J curve still looks flat only because the scale broke the chart. Twelve weeks is how you get in front of it instead of under it.

Let's talk

If you want a copy of the data-dependency checklist behind this plan, or a view on what your firm's twelve weeks would look like, drop me a line at dave.howard@wealthtechpros.com.

The first three weeks cost you an audit and a decision. Standing still costs more, and it compounds.

My thanks to Paul Miller, a proper-job data-warehouse architect steeped in wealth-management expertise, for his practitioner insights.