PAPER A
Discovery Sprint
2 weeks
from£4,500
- Map one decision-heavy workflow
- Design the agent team: roles, rules, and what humans decide
- A working prototype session on your own question
AI agent teams for UK businesses
Your hard decisions, researched, debated and voted on by a team of AI agents.
Every answer has to win the vote, and you get signed minutes. Your people decide anything that changes the rules.
Council · week 43
Sorting this week’s open questions
Commercial
42
4 for, 0 against, 1 objectAmended
18
5 for, 0 against, 0 objectCarried
31
3 for, 0 against, 2 objectTo sign-off
16
1 for, 3 against, 1 objectRejected
Operations
36
4 for, 1 against, 0 objectCarried
22
1 for, 3 against, 1 objectRejected
27
4 for, 0 against, 1 objectAmended
24
5 for, 0 against, 0 objectCarried
Illustration · every card is a finding, source or objection; every circle is a vote. Click a decision to read its vote.
Founder-led. UK based. We also build and run our own AI product, HermanWa, deployed across UK, UAE and GCC property portfolios. Every session leaves minutes, sources, a tamper-evident log and a cost line.
01The problem
One chatbot gives you one confident answer and no record.
A single assistant will answer anything in the same assured tone. It does not tell you what it checked, who disagreed or what it cost. When the decision is questioned in three months, there is nothing to show.
A single chatbot
- One voice, one view, no challenge
- Sources optional, often missing
- Certainty is the same for every answer
- No minutes, no log, no cost per question
A governed agent team
- Specialists with set roles and memory
- Every finding carries its source
- Objections must carry a fix, then a vote
- Minutes, a hash-chained log and a cost line for every session
02How it works
Your agents work under standing orders.
We write the rules of the room with you before any agent runs. These are the defaults we start from.
- Specialists, with roles
Each agent owns one area, such as finance or regulation, and keeps its own memory between sessions.
- Evidence carries a source
Findings cite a document, dataset or report. Uncited claims are marked as opinion in the minutes.
- Objections carry a fix
An agent can block a proposal only by saying what would make it acceptable.
- The council votes
For, against, or object. The tally and each agent's reason go on the record.
- The chair decides routine calls
A chair agent rules on the vote within limits you set, and folds in amendments.
- People decide rule changes
Anything that changes a policy, price or rule waits for a named human to sign off.
- Everything is recorded
Minutes, a tamper-evident log where each entry hashes the one before, and spend per session.
- Lab mode for hard questions
A literature survey with live sources, position papers, assigned critics, a novelty check, then experiments whose pass rules are written before the data.
03Sneak view
Watch a council session from question to minutes.
A staged replay of a short session on an everyday business question. The agents, sources and figures are illustrative.
Close up: one vote
Every specialist votes with a reason: for, against, or object. An objection only counts if it comes with a fix, so “no” on its own is not allowed.
The chair folds the fixes in and records the result. Here the regulatory agent's fix is adopted and the motion carries as amended.
Illustrative session
“Should we launch our product in Germany in Q3?”
Session flow
- 1QuestionTabled with budget and rules
- 2Specialists investigateFindings with sources
- 3Debate and objectionsEach objection carries a fix
- 4VoteFor, against, object
- 5Chair decidesWithin set limits
- 6Minutes and audit trailHash-chained log, cost
- HHuman sign-offOnly where a rule changes
The highlighted step follows the replay.
04What you get
Decisions you can defend, with the paperwork attached.
Governance
Written roles, standing orders and limits on what the chair may decide alone. You choose which decisions always come to a person.
e.g. pricing, policy, anything customer-facingMinutes
Every session ends with decisions, actions with owners, dissent and sources, in plain English your team can file.
decisions · actions · objections · sourcesCost visibility
Spend per session and per agent, with a cap set before the session starts. Sessions stop at the cap.
e.g. £3.80 for the session aboveHonest uncertainty
Agents state confidence and what would change their mind. On the retainer we track how often their forecasts came true.
confidence · forecast accuracy over timeAn audit trail
A tamper-evident log: each entry carries a hash of the one before, so an edited record shows.
entry 41 → sha256(entry 40 + content)Your stack, your data
Built in your cloud account or on your servers, with the model providers you have approved.
handover, training and documentation included05Your data
Your data stays yours. Here is how, path by path.
A chatbot has one way to leak your data: the model provider. An agent team has five, because agents search, fetch, remember and write. We design a control for each one before any agent runs.
| # | Where data could leave | What could go wrong | Our control |
|---|---|---|---|
| 1 | The model provider | Your prompts are kept, or used to train future models | Commercial terms with no training on your data, zero data retention where the provider offers it, UK or EU regions, or models run inside your own cloud account |
| 2 | The agents' tools | A web search sends its query to a search provider; a fetched page can hide an instruction telling an agent to send data out | Agents on confidential work get no web access, or generalised queries only; an allowlist of sites; a named person approves anything that leaves |
| 3 | Memory | One client's facts surface in another client's session | A separate deployment, memory and log for every client. Nothing is shared |
| 4 | The records | Minutes and audit logs hold the sensitive content | Kept in your environment, encrypted, access-controlled, deleted on the retention period you set |
| 5 | People and outputs | A case study, demo or shared link reveals too much | Anonymised by default, nothing public without your written sign-off, private sharing |
TIER 1
Standard
A commercial model service with no training on your data, UK or EU region where available, the agents built in your cloud account.
TIER 2
Private cloud
The models run through your own AWS, Azure or Google account, so your data stays in your tenancy under contracts you hold.
TIER 3
Fully private
Open-weight models on your own hardware with no outside calls. The most private option, at some cost in capability.
What we promise: your data stays in your environment, is never used to train a model, nothing leaves without a person approving it, and every action is logged so you can check.We will not promise zero risk. We will show you each control working before your data goes in, and sign a data processing agreement under UK GDPR with the sub-processor list.
06Case study
Nine specialists, one night, a defensible research plan.
We run our own advanced simulation and machine-learning R&D programme with a nine-specialist AI research team. Here is what one piece of that work produced.
Our own programme · numbers from our own records
One deep research session, run overnight, for about 1.1M model tokens.
8research frontiers surveyed
37citations verified live
~21kwords across 8 position papers
24peer critiques from assigned critics
4new method lines, each with a registered first experiment
9specialist agents in the council
14.7%of compute found doing unused work, then removed
1.67×faster, with byte-identical results
- Speed to a decision. The survey, papers, critiques and a novelty re-check finished in one night. We started the next day with experiments already specified.
- The review caught itself. The process found a wrong assumption in its own debate and corrected the record, with the correction minuted.
- Amendments, folded in. A three-agent review of a design returned “approve with amendments”. Every amendment went into the design before build.
- Cost visibility. A waste audit found 14.7% of compute producing output nothing used. Removing it made the measured run 1.67× faster with byte-identical results.
07Why us
We run an agent team on our own work before we build one for yours.
Nothing on this page was designed on a whiteboard. The standing orders, the votes, the minutes and the cost lines are how we run our own research and development, every day. When something went wrong, we changed the rules and minuted why.
Our own team, in daily use
A nine-specialist agent council runs our simulation and machine-learning R&D: research, statistics, engineering and review roles, each with its own memory, with minutes and a cost line for every session.
the case study above is our own programmeRules learnt the hard way
An independent critic agent reviews every important result before it goes on the record. Expensive runs wait for a full-length trial first. When the team got something wrong, the correction was minuted, not quietly fixed.
each rule came from a real mistakeA live AI product
We build and run HermanWa, our own AI product, deployed across UK, UAE and GCC property portfolios. We know what it takes to keep AI working in front of paying customers, not just in a demo.
HermanWa is a trading name of Danai Data UK LtdAgent teams for other companies
We have designed and deployed governed agent teams for businesses in real estate, tourism, management consulting, manufacturing and technology, each in the client's own environment, with the controls on this page.
client names withheld under NDATHE FOUNDER
Herman Wa
Herman has spent seventeen years building data systems that have to stand up to auditors, investors and boards: leading sustainability and data in FTSE-listed real estate, global tourism and NHS healthcare, across multi-billion-pound portfolios, with independent assurance year after year.
Herman now designs and deploys governed AI agent teams for real estate, tourism, management consulting, manufacturing and technology businesses, and builds them hands-on: the agents, the model integration, the data pipelines and the controls. Every client team runs in the client's own environment, under the rules, minutes and audit trail on this page.
Herman also runs a nine-specialist agent team on a demanding in-house simulation and machine-learning R&D programme; its results are the case study above.
Plenty of people can build an AI agent. Few have spent years answering to auditors. Herman does both, and that is what makes an agent team safe to rely on.
MEng Architecture & Environmental Design, University of Nottingham · PG Diploma in Surveying, University of Reading · GRI Certified Sustainability Professional · Microsoft Certified Power BI
Start small: every engagement begins with a fixed-price Discovery Sprint.It ends with a working session on one of your own questions, so you judge the result before committing to anything larger.
08Packages
Start with one decision. Grow from there.
PAPER B
Agent Team Build
4 to 6 weeks
from£18,000
- The team built in your stack
- The council, with recorded minutes
- Cost controls and session caps
- The tamper-evident audit trail
- Handover and training for your team
PAPER C
Run and Improve
Monthly retainer
from£3,500/month
- We operate and tune the team
- Deep research sessions on your questions
- Monthly report: decisions made, forecast accuracy, spend
Prices exclude VAT. Model usage is billed at cost and shown per session.
09Who it is for
Teams in mid-sized companies that make judgement calls on evidence.
Strategy
Market entry, pricing moves and portfolio choices, argued from several angles.
Research
Literature surveys, competitor scans and experiment plans with pass rules set first.
Finance
Investment cases and forecasts with stated confidence and a record of who disagreed.
Legal-ops
Policy and contract review with every point sourced and every change signed off.
Product
Roadmap trade-offs weighed by agents for customers, engineering and cost.
10Questions
Frequently asked
Where does our data go?
It stays in your environment. We build in your cloud account or on your servers, and agents read only the sources you connect. Model calls go to providers you have approved, under business terms with no training on your data. The five controls are set out in Your data.
What qualifies you to build this?
We have deployed governed agent teams for real estate, tourism, management consulting, manufacturing and technology businesses; client names stay under NDA. We run one on our own R&D every day, and we build and run HermanWa, an AI building-management platform deployed across UK, UAE and GCC portfolios. The founder spent seventeen years in audited data roles in FTSE-listed real estate, global tourism and NHS healthcare. See Why us.
Which models do you use?
We pick per role. A finance agent may need a stronger reasoning model than a note-taker. We work with the major commercial model providers and can use models hosted in your own cloud. You see the cost of each choice before it goes live.
Do the agents make decisions on their own?
Only the routine ones you allow. The chair agent rules within limits you write down. Anything that changes a rule, a price or a policy waits for a named person to sign off, and that sign-off is logged.
What does it cost to run?
A short council session like the one above typically costs a few pounds in model usage. A deep research session in lab mode costs more, depending on how many specialists and sources it uses; we quote it and set a cap before it starts. The minutes show the actual spend.
What if the agents are wrong?
They will be sometimes. That is why findings carry sources, agents state confidence, objections are recorded and people sign off rule changes. On the retainer we report how often the agents' forecasts came true.
How long until we see something working?
The Discovery Sprint ends with a working session on one of your own questions, two weeks after we start.
11From the blog
Notes on running agent teams under rules.
What is an AI agent team? The plain-English guide for UK business leaders (2026)
Specialist AI agents that research, argue under rules, vote and hand you minutes. Here is what that means, how it differs from a chatbot, and when it is worth having.
Read the post
How much does an AI agent team cost in the UK? Build prices, running costs and the 14.7% waste we found
The two costs that matter are building the team and running its sessions. Here is how both work, why every session should have a cap, and what our packages cost.
Read the post
We ran 9 AI agents overnight on our own R&D. Here is what they produced, and the 6 rules we learnt the hard way
Before building agent teams for clients, we run one on our own simulation and machine-learning research. Here is what one overnight session produced, and the rules we added after things went wrong.
Read the post12Contact
Book a discovery call.
Thirty minutes with the founder. Bring one decision your team keeps revisiting. We will tell you plainly whether an agent team would help.
dan@danaidata.com