Public sector

Automation a public body can stand behind

Public bodies are asked to show their working: to auditors, to the people affected, sometimes to a court. We build automation, and the evidence around it, with that reader in mind from the first day.

Questions you should expect to be asked

Each of these has a specific answer in what we deliver, based on the trust framework.

The questionWhat we provide
Why is this being automated at all?A warrant: the evidence that it is lawful, tested and defensible. It opens with the purpose and business case, the lawful basis, the accountable owner, the risks accepted and the standard it must meet before it is relied on.
Who made the decision?Decisions about meaning, eligibility or consequence are made by a named person. The record shows who, and when.
How do you know it works on our data?A test against a golden set produced by your own experts, with agreement measured using proper statistics and every figure reported with a confidence interval. The sample size and the limits are stated.
Can you show your working for this case?A full audit trail from raw data to output, so any figure or quote can be traced to the source behind it.
Where does the data go, and who sees it?Agreed and documented before work starts: processing location, access, retention and deletion, and whether any AI service may keep or learn from it.
What happens when it is wrong, or unsure?Items that do not fit, or where independent runs disagree, are left unresolved and shown to a person. The limits are written on the trust sheet.
What if we change supplier?The tests, the records and the documentation belong to you and are portable. Nothing depends on a platform of ours, because there is not one.

How we would run an engagement

  1. Decide. Through a structured interview with the people accountable, agree the purpose and business case, the lawful basis, the accountable owner, the risks accepted, and what good enough means. This opens the warrant. The outcome may be that the process should not be automated.
  2. Design. Map the process and separate what is deterministic, what is routine judgement and what is a decision about meaning. Constrain the AI, and decide what should never be automated.
  3. Build the golden set with your experts. Their own work becomes the standard the automation is tested against.
  4. Build, attack and validate. Test it adversarially, then improve it until it meets the threshold in the warrant. Write down the result, with confidence intervals, and what it does not cover.
  5. Run with review and a record. People decide at the agreed points, and everything is recorded as a by-product of the work.
  6. Hand over. The warrant, the trust sheet, the records, the evidence and the runbook, so your team can operate and challenge it without us.

How this fits the paperwork you already produce

We design the record so that it can feed the documents public bodies already have to produce: a data protection impact assessment under UK GDPR and the Data Protection Act 2018, an entry in the government's Algorithmic Transparency Recording Standard, supplier assurance answers, an internal register of AI use. Where a service reaches people in the EU, the EU AI Act may apply too.

Where an automated process could produce a decision about a person, whether the UK GDPR rules on decisions made solely by automated means apply is answered in the warrant at the start, not discovered later.

We are not a law firm and do not give legal advice. What we do is make sure the facts those documents need exist, and are accurate.

What we will not do

  • Automate decisions about an individual's entitlements without a person deciding.
  • Quote performance figures that have not been measured, or describe as validated something that is not.
  • Claim certifications we do not hold.
  • Present automation as a way to remove the people accountable for the outcome.

Where to start

A good first step is small: one process, and a written view of what could be automated, what should not be, and what evidence you would need before relying on it. That gives you something to take to your governance board whether or not you go further.