Phase 2: department brains
Phase 2 gives each department a brain: a knowledge layer that answers questions across the team's own data, with sources. Agents that analyse, enrich and monitor then read from it. This is where bespoke build enters the roadmap.
A department brain answers questions across one department's own documents, records and data, and shows its sources. It is built by retrieving the relevant records before the AI writes an answer. Phase 2 also adds agents that read from the brain to analyse, enrich and monitor, while people keep every decision.
What is a department brain, and how do we build one?
A department brain answers questions across one department's own documents, records and data, and shows the sources behind each answer. You build one by organising that data, connecting it to a system that finds the relevant records for each question and letting the AI answer only from what it found.
The department brain is the centre of Phase 2 of the AI-Native Roadmap. It comes after Phase 1, because it depends on a team that already uses AI daily and data that Preparation has put in order.
A department brain is what turns AI that knows the world into AI that knows your organisation. It moves every department closer to the roadmap's destination: An AI-native business is one where AI assistants, department brains and agents do the repetitive work in every team, with people keeping the judgement.
What can a department brain answer that an assistant cannot?
A department brain answers questions about your own organisation: what a customer was promised, which version of a policy applies, what happened the last time a situation came up. A general assistant can only guess at these, and it guesses fluently.
That fluency is the problem a brain solves. A general model writes an equally confident answer whether or not it has the facts. A brain answers from records it has just retrieved and shows them, which makes a wrong answer far easier to spot.
- Sales: "What did we quote this client last year, and what did they push back on?"
- HR at a school: "What does our staff handbook say about leave for a family emergency?"
- Fundraising at a charity: "Which trusts funded youth work in the last three years, and what did we report back?"
- Operations at a sports body: "Which competition rule applies to a player registered with two clubs?"
Few organisations have reached this point. The Office for National Statistics (opens in a new tab) found that among UK businesses with 10 or more employees that use AI, only 10% use it extensively. McKinsey's 2025 global survey (opens in a new tab) found nearly two-thirds of organisations have not yet begun scaling AI across the enterprise.
What data work does a department brain need?
A brain is only as good as the data it reads, so most of the work in Phase 2 is data work. Each source is identified, tagged, kept current and audited on a cycle, so the brain never answers from last year's version.
- Static sources: policies, handbooks, templates, past proposals and reports. They change rarely, so each needs an owner and a review date.
- Dynamic sources: the customer system, the finance system, shared mailboxes and meeting notes. They change daily, so they connect live or refresh on a schedule.
- Tagging: each document marked with its type, date, owner and who may see it, so the brain can filter before it answers.
- Audit cycles: a monthly check that retired documents have gone and new ones have arrived, plus spot checks of the brain's answers against the sources.
Access controls matter as much as accuracy. IBM's 2025 Cost of a Data Breach report (opens in a new tab), which studied 600 organisations that had suffered a breach, found that 13% reported breaches of AI models or applications. Of those, 97% reported not having AI access controls in place.
Personal data needs particular care. The brain should let each person see only what they could already see, and the questions in the AI governance and data protection work, such as the lawful basis and the supplier's terms, are settled before the build starts.
How is a department brain built?
A department brain is built in two parts: a search layer that finds the passages relevant to a question, and a language model that writes the answer from those passages and cites them. The model is told to answer only from what was found.
The technique is called retrieval-augmented generation, which means the AI looks the answer up before writing it rather than relying on memory. It was set out in a 2020 research paper by Lewis and colleagues (opens in a new tab) and now underpins most systems that answer from a company's own documents.
The search layer usually stores each passage in a vector database, a store that finds text by meaning rather than by exact words. So a question about "time off for a bereavement" still finds the policy headed "compassionate leave".
Platform or bespoke?
Start with a platform where one can do the job. Several business systems now answer from a defined set of company documents, and for a single team with tidy files that may be enough.
This is where bespoke build enters the roadmap, at the point where a platform cannot do the job: the brain must combine several systems, respect detailed permissions or fit a process no product supports. Bespoke AI solutions explains how Hartz AI scopes that work, and deployment and integration covers connecting it to the systems you run.
Which agents read from the department brain?
Phase 2 also adds the first AI agents: software that pursues a goal by choosing its own steps within limits a person sets. They read from the department brain, which is what makes their work reliable enough to use.
- Analyse: an agent that reviews each week's support tickets and reports the five most common causes.
- Enrich: an agent that completes a new donor or customer record from what the organisation already knows.
- Monitor: an agent that watches for contracts, grants or memberships nearing renewal and drafts the next step.
Each agent has one job and a narrow set of permissions. Agents that coordinate with each other, under an orchestrator agent, belong to Phase 3.
What does a person still decide in Phase 2?
People decide everything that matters: whether an answer is right, what to do about it and anything that reaches a customer, a member, a funder or a decision. The brain and its agents prepare; people judge.
A synagogue office is a useful picture. Its brain can tell the administrator which members have not yet renewed and what was agreed with each family last year. Whether to call, what to say and whether to waive a fee stay with the administrator and the rabbi.
A department brain makes people faster at judgement, not unnecessary to it.
When is a department ready for Phase 3?
A department is ready for Phase 3 when its brain passes its monthly accuracy checks, its first agents run reliably with few corrections and the team trusts both enough to hand over longer chains of work.
The signal to watch is people asking for the next step: "could it draft the reply as well?" or "could it book the follow-up too?" When single agents start handing work to each other by hand, the department needs orchestration.
Common questions
Frequently Asked Questions
Plan a brain for one department
A discovery call maps the data a department brain would read, the permissions it needs and whether a platform can do the job.