DataFab / Utilities / Consumer-redress remediation
Governed utility · remediation programmes at population scale
Population-scale redress, defensible on the day it ships.
A redress programme is a data problem wearing a legal deadline. The affected population has to be identified across systems that disagree, the calculation has to be deterministic and reproducible, the outreach has to happen at volume, and the whole thing has to stand up on delivery rather than be reconstructed under pressure afterwards.
The brief
“A defined population, a fixed deadline, a client environment we cannot move data out of, and a redress calculation that has to be right to the penny and explainable per customer — delivered with people we have to recruit, train and then release.”
Remediation is priced and staffed as a people problem because that is how it has always been executed: a bench assembled against a deadline, working case by case through systems that disagree, producing an evidence pack at the end from what the bench happened to record along the way.
The work does not actually need that shape. Identification is a resolution problem. Eligibility is a versioned rule. Calculation is deterministic and should never involve model discretion. Outreach is volume with an audit trail. What genuinely needs judgement is a narrow slice — and that slice deserves the whole of your senior time rather than a fraction of it.
Closed books count
Origination, servicing, collections, complaints, archive and whatever the last acquisition left behind. The population is spread across all of it, and the oldest systems hold the hardest cases.
Fixed, and public
Capacity cannot be added late without the quality of the pack falling. Throughput has to be designed in from the first week, not discovered in the last.
Per customer, not per programme
The question is not whether the methodology was sound. It is why this customer was included, excluded, or paid this amount — answered thousands of times over.
Data cannot move
Customer records stay inside the client’s boundary. Any approach that starts by staging a copy into a remediation database starts by creating a new governance problem.
End to end
One governed pipeline, read-only across the estate.
Nothing is migrated into a remediation database. The fabric reads the systems already running, resolves the customer across them, and runs the programme over that resolved population — so the audit trail and the working data are the same thing.
Cohorting & eligibility
Inclusion is a versioned decision, per customer.
Cohorts are built over the resolved population rather than over one system’s view of it — grouped by product, period, commission model, disclosure state and whatever else the methodology turns on. Every customer carries the rule version that put them in scope and the fields it read.
When the methodology moves — and it always moves — the change is a new rule version, not a re-run that quietly rewrites history. Both versions remain replayable, and the population delta between them is itself a reviewable artefact.
| Question the reviewer asks | Where the answer comes from |
|---|---|
| Why is this customer in scope? | Rule version, fields read, values at the time |
| Why is this one excluded? | Exclusion rule, evidence, approving authority |
| Are these two the same customer? | Resolution decision, confidence, sources merged, reversible |
| What changed since the last cut? | Population delta between two rule versions |
| Did the source say that at the time? | Snapshot of the cited record as it stood |
One customer holding three agreements across two acquired books is one customer, three cases — or three customers and a reconciliation exercise, depending entirely on whether the estate was resolved before the cohorts were cut.
Calculation
Deterministic, by design.
The automation dial is set per step, and for redress arithmetic it is set to the bottom: fixed steps, reproducible, no model discretion anywhere in the flow. Extraction of a document may run agentic; the number a customer is paid never does.
Deterministic
Fixed steps, reproducible, no model discretion in the flow.
- Where it fits
- Fee checks, redress calculations, statutory schedules, fixed procedures.
- Model discretion
- None. The rule decides, every time, identically.
- What it leaves behind
- Reproducible to the penny. Replayable under the rule version in force.
Process-governed
A defined business process routes the work; rules drive the flow.
- Where it fits
- Regulated workflows, service-level agreements, sign-off chains.
- Model discretion
- Routing only, inside a declared process.
- What it leaves behind
- The process is the audit trail. Every branch taken is recorded.
Human in the middle
Agencies do the work; a person approves at each gate.
- Where it fits
- Coverage decisions, conflicts, filings, anything contestable.
- Model discretion
- Prepare and propose. Never dispose.
- What it leaves behind
- The gate blocks until approved, and the approver and their basis are logged.
Self-organising
Agencies route and adapt within policy and schema bounds.
- Where it fits
- Triage, enrichment, discovery, evidence gathering.
- Model discretion
- Chooses its own path — inside the bounds you set.
- What it leaves behind
- Every path taken is recorded; the bounds themselves are versioned.
Fully agentic
Autonomous reasoning inside the sandbox and the schema.
- Where it fits
- Extraction, summarisation, low-stakes work.
- Model discretion
- Full, within an isolated runtime and a bound schema.
- What it leaves behind
- Post-hoc audit. It cannot invent entities outside the schema.
And it is set per step — one written claim, eight decisions, five different levels
Outreach
AutoFab runs contact at volume.
Contact is the part that quietly consumes a programme: attempts, channels, non-responses, returned mail, re-contacts, vulnerable-customer handling and the record of all of it. AutoFab runs it as governed work, with every attempt and every response attached to the case rather than to a spreadsheet somebody maintains.
When a customer responds, the response re-enters through intake and the case re-evaluates automatically — eligibility reopened, calculation refreshed, no manual re-keying and no queue of envelopes waiting for somebody to type them in.
Every attempt
Channel, timestamp, outcome and the template version used — per customer, on the case.
Every non-response
Silence is a state with a policy attached, not a gap in the file.
Vulnerability handling
Flagged cases route to a person under a defined policy rather than continuing down the automated path.
Response re-enters
Intake, re-evaluation and recalculation without a human retyping anything.
The evidence pack
The file and its evidence are the same artefact.
Packs assembled at the end of a programme are reconstructions, and they read like it. Here the record is produced as the work happens, per case, and the pack is a view over it rather than a project of its own.
Delivery economics
Price the outcome, not the bench.
For a delivery partner this is the part that changes the business, not just the programme. When governed agencies do the work and a person sits on the gate, pricing moves from per full-time equivalent to per case — and because the fabric compounds, unit cost falls as volume grows rather than staying flat.
Today · priced per person
Margin competed away, head by head.
- Cost driven by analyst time
- Unit cost flat as volume grows
- More throughput means more recruitment
- Quality varies with the bench you could hire
- Nothing carries to the next programme except people
Governed · priced per case
Margin designed in, with a floor.
- Agencies do the work; the analyst sits on the gate
- Unit cost falls at volume as the fabric compounds
- Throughput scales without recruitment
- Quality is a property of the pipeline, not the intake cohort
- The encoded methodology is an asset you keep and redeploy
Illustrative shape of the economics, not a quote. Real figures are set per engagement against case mix, volumes and clearance policy.
A programme ends and the bench disperses. The encoded methodology does not: the cohorting logic, the eligibility rules, the calculation, the outreach flows and the evidence structure are governed, versioned objects your firm owns and points at the next book, the next client, the next jurisdiction.
Where it runs
Inside the client’s boundary.
A delivery partner rarely controls the environment. The data plane runs inside the client’s cloud, VPC or data centre under the client’s keys; the managed control plane holds policy, metadata and orchestration and never holds raw customer data. Your people operate the programme; the client’s controls stay the client’s.
Next step
Bring the population and the deadline.
The first test is the cohort nobody can cleanly identify, across the books nobody wants to migrate. Resolution makes the true population visible before a single case is worked.