Building on the Right Layer: What Agentic AI Actually Needs from Government Data
Most public sector conversations about agentic AI focus on governance, ethics, and risk. Necessary, yes. But there's a more immediate blocker: our data foundations aren't ready.
The promise of systems like the GOV.UK Agentic AI Companion depends on fast, programmatic access to high-quality, interoperable data. Today, that's rare. The UK's own reviews suggest most government data is still siloed and hard to join up. Global benchmarks show the same pattern: pilots start, then stall, because the infrastructure isn't there to support production use.
The five rungs of agent-ready data
Think of "agent-ready" as a ladder. Most teams are stuck on Rung 1 or 2. The leap to Rung 3 is where pilots fail. Here's what needs to be true before an agent can actually act for citizens, safely and at scale.
Rung 1: Catalogued
What it means: Datasets are discoverable, described, and locatable.
What the agent needs: A clear picture of what exists and where it lives.
The gap: Data hiding in spreadsheets, inboxes, and undocumented systems is invisible to any agent.
Rung 2: Quality-assured
What it means: Accuracy, completeness, and timeliness standards are defined and measured.
What the agent needs: Reliable inputs to avoid flawed decisions.
The gap: Agents won't pause to sanity-check every field. Old or inconsistent data drives bad advice.
Rung 3: Accessible
What it means: API-first, machine-to-machine access by default.
What the agent needs: Data in seconds, without forms, emails, or manual extracts.
The gap: Human-mediated access breaks autonomy. If a person has to fetch it, the agent can't use it.
GDS API standards set a clear bar for this step.
Rung 4: Observable
What it means: Provenance, audit trails, and decision tracing are built in.
What the agent needs: A way to reconstruct which data it used, what it inferred, and why it acted.
The gap: Without lineage, you can't explain outcomes or correct them.
Rung 5: Interoperable
What it means: Shared data standards and semantics across departments and partners.
What the agent needs: Seamless joins across datasets like housing, employment, benefits, and residency.
The gap: Constant bespoke translation is slow and costly. It doesn't scale to cross-government use cases.
Start with common UK standards (for example UPRN/USRN, ODS codes) and the guidance from the Data Standards Authority.
Accountability gaps you can't ignore
- Accountability in error: If advice was wrong because the dataset was outdated (a Rung 2 failure), who is responsible-the data owner, the deployer, or the vendor?
- Oversight vs. autonomy: Agents act on their own, but citizens expect human accountability. Where is the line drawn and documented?
- Rollback strategy: If an agent gave incorrect advice for weeks, how do you find affected people, fix outcomes, and prevent recurrence?
A practical 90-day plan to move from pilots to production
- Week 1-2: Inventory what you've got. Produce a live catalogue of the top 20 datasets agents would rely on. Note owners, SLAs, and current access methods.
- Week 2-4: Set SLOs for data quality. Define accuracy, completeness, and freshness targets with automated checks. Publish results to a shared dashboard.
- Week 3-6: Go API-first for the top tasks. Expose machine-readable endpoints for the 5-10 most used data items. Apply the GDS API standards, with auth, rate limits, and versioning.
- Week 4-8: Add observability. Capture dataset versions, query logs, and decision traces. Make them queryable for audits and user complaints.
- Week 6-10: Standardise semantics. Map to shared identifiers (e.g., UPRN/USRN, ODS) and adopt DSA guidance for priority domains.
- Week 8-12: Prove accountability. Define decision ownership, sign-off criteria for autonomous actions, and a named data steward per dataset.
- Week 10-12: Run an "agent fire drill." Simulate a bad outcome. Trace affected users, issue corrections, and document the playbook.
- Upskill the team. Short, focused training on data quality engineering, API product management, and auditing can shave months off delivery. If helpful, see our curated options for public-sector roles: AI courses by job.
Constructive optimism
The work here isn't new. Catalogues, quality, APIs, audit trails, and standards have been on government roadmaps for years. Agentic AI just raises the stakes and exposes where that work stalled.
The good news: agents don't need a brand-new data stack. They need the existing strategy executed well. Treat this as accelerated foundation-building, and you move beyond perpetual pilots to safe, scalable delivery.
Interested?
If this topic matters to you, register to attend our Think Data for Government conference, where digital leaders tackle the most pressing data issues facing government today. Register Now.
Liked this article? Subscribe to our newsletter to receive the latest digital transformation news directly to your inbox.
Your membership also unlocks:
