Case study · Helios Logistics

Route to Terraform

One Function — 'Generate customer invoice' — across all six rungs of the adoption ladder. 21 months, 5 decision records, 2 emergent roles, 6 → 0 specialists, 18 → 0 days cycle time. Click each rung to walk through what changed and why.

Subject Function

Generate customer invoice

Roll up a customer's shipments for the month into a single invoice. Apply the right rate cards, surcharges, and contracted discounts. Reconcile with payments.

Billing (Finance) · owned by Tom Whitfield, Head of Billing
Pre-platform baseline
Cycle time18 days
Error rate4.2%
Team size6 specialists
Monthly toil cost£40k
Span
21 mo
Total spend
£2.26m
Annual toil saved
£480k
Headcount delta
-3
2 emergent roles
Customer NPS
+28
Rung 0 · Pre-Q1 2025
Baseline · ~years
ManualBaseline

Six specialists, eighteen-day cycle, four percent errors

This is where Helios started. Billing specialists pulled shipment data out of the TMS, applied rate cards by hand in spreadsheets, drafted invoices in DocuFlow, and emailed them. End-of-month was a sprint; mid-month was preparation for the sprint.

Cycle time
18d
Error rate
4.2%
Team size
6
Monthly toil
£40k
Trigger
Objective
(none — baseline)
No active Objective. Adoption Platform was being used to model the org and measure where AI uplift would be most valuable.
Scenario chosen
(no scenario — manual baseline)
The manual process was documented and instrumented as the basis for measuring future improvement.
Workforce impact

6 Billing Specialists (4 senior, 2 junior). Stable team; high turnover risk in juniors due to the repetitive nature of the work.

Assessment
Baseline benchmark — Q1 2025
pass
process documented: passmetrics instrumented: passdata-quality baseline: marginal

Shipment-manifest quality was already flagged as a constraint on any future automation.

We knew billing was eating the team. The question was whether we'd get more value from making the team better at it, or from making the work itself smaller.

Tom Whitfield, Head of Billing
Rung 1 · Q1 → Q2 2025
9 weeks build · 2 months in production
AssistedShipped

Drafting copilot lands; specialists spend less time typing

An AI copilot drafts each invoice from the shipment manifest. The specialist still reviews every line, applies any non-standard adjustments, and presses Send. Human is firmly in the driver's seat. The copilot writes maybe 70% of the words.

Cycle time
14d
-22%
Error rate
3.8%
-0.4 pp
Team size
6
+0
Monthly toil
£30k
−£10k/mo saved
Trigger
Objective
Reduce billing toil by 25%
Customer Success was hearing that finance teams at our customers were tired of late, error-prone invoices. CFO sponsored a low-blast-radius first move.
Scenario chosen
Drafting copilot (assisted)
Bolt a drafting agent next to the specialist's existing workflow. No process change. No role change. Train on 12 months of past invoices.
£86k·2.1 mo·high
Alternatives considered
  • End-to-end automation (autonomous)Too aggressive for a first move. Workforce impact unclear; data quality unproven.
  • Templating-only (no AI)Wouldn't move the needle on cycle time. Past attempts had stalled.
Workforce impact

No role changes. All 6 specialists trained on the copilot UI in two half-day sessions. New skill added to their profiles: 'AI-assisted drafting'.

Assessment
Drafting copilot — release gate
pass
brief conformance: passcorrectness: passoutcome: passworkforce-intent: pass

Passed first-time. Specialists actively asked for the copilot to be expanded.

The first version was honestly nothing fancy. But once the team saw the AI didn't make them dumber, they got curious. That curiosity was worth more than the four-day cycle-time win.

Tom Whitfield, Head of Billing
Audit reference: dr-bill-2025-q1
Rung 2 · Q3 2025
11 weeks build · 1 quarter in production
AugmentedShipped

Multi-doc reasoning takes the bulk; humans handle exceptions

The agent now reads shipment manifests, rate cards, contract addenda, and prior disputes — produces the invoice, including all line items, surcharges, and adjustments. The specialist no longer 'reviews every line' — they review summaries and exceptions only. Routine invoices flow through untouched.

Cycle time
9d
-50%
Error rate
2.1%
-2.1 pp
Team size
5
-1
Monthly toil
£18k
−£22k/mo saved
Trigger
Objective
Continue billing toil reduction · close gap on rate-card freshness
First wave validated the pattern. Drift signal showed the copilot was inconsistent on contract addenda — not an error, but exposed an opportunity to expand its remit.
Scenario chosen
Multi-doc reasoning (augmented)
Expand the copilot's context to include rate cards, contract addenda, and prior dispute history. Adopt an exception-only review pattern.
£142k·2.6 mo·med
Alternatives considered
  • Stay at assisted, do data-quality work firstWould have been the safer call. CFO chose the more ambitious path because the team was confident.
Workforce impact

1 junior specialist position not back-filled after natural attrition. Team is now 5 — two seniors developed deeper exception-triage muscle.

Emergent roles
  • Exception Triager (Billing) — predicted, not yet promoted
Assessment
Multi-doc reasoning — release gate
marginal
brief conformance: passcorrectness: marginaloutcome: passregulatory: pass

Marginal on correctness because rate-card refresh lag caused 4% of invoices to use stale prices in the first two weeks. Mitigated with a freshness-check + auto-pause.

Drift event during this stage
medium
Stale rate-card detection — 4% of invoices used prices ≥7 days out of date for 12 days

Resolution:Added a freshness-check at invoice generation; if rate-card timestamp older than 72h, auto-pause and notify Tom. AdoptionState briefly auto-demoted to assisted; restored after fix.

The marginal assessment was the most useful thing that happened that quarter. It was a real signal, the platform forced us to address it, and the rate-card freshness fix paid off everywhere — not just in billing.

Helena Vargas, CFO
Audit reference: dr-bill-2025-q3
Rung 3 · Q4 2025 → Q1 2026
16 weeks build · 4 months in production
Supervised-AutonomousShipped

Self-disputing flow — agent handles reconciliation, humans approve outliers

The agent doesn't just generate the invoice — it watches the customer's payment flow, reconciles automatically, and opens dispute drafts itself when something looks wrong. The specialist now approves outliers only: cases where the agent flagged uncertainty or where customer-facing language matters.

Cycle time
4d
-78%
Error rate
0.9%
-3.3 pp
Team size
4
-2
Monthly toil
£8k
−£32k/mo saved
Trigger
Objective
Reduce customer dispute volume by 40%
Customer Success requested. Disputes were eating 30% of CSM time and tarnishing renewal conversations.
Scenario chosen
Self-disputing copilot (supervised-autonomous)
Agent generates invoice, reconciles payment, drafts dispute response if customer queries. Human approves the dispute response and any uncertain reconciliation.
£268k·3.7 mo·med
Alternatives considered
  • Build a separate dispute-only agentWould have created a seam that loses context. Better to extend the existing agent's remit.
  • Outsource disputes to a BPOConsidered briefly; would have killed the workforce-evolution story.
Workforce impact

2 specialists transitioned to a new emergent role: Customer Outcome Steward. They own the customer-facing language, edge cases, and the agent's training feedback loop.

Pathways
  • Noah Kowalski · Billing SpecialistCustomer Outcome Steward · completed
  • Ines Marchetti · Billing SpecialistCustomer Outcome Steward · completed
Emergent roles
  • Customer Outcome Steward — approved, filled
Assessment
Self-disputing copilot — release gate
pass
brief conformance: passcorrectness: passoutcome: passworkforce-intent: passregulatory: pass

First-time pass. Drift thresholds set tighter than usual due to customer-facing exposure.

Drift event during this stage
low
Dispute-tone drift — agent's tone became more formal than the brand voice over 4 weeks

Resolution:Customer Outcome Steward feedback loop caught it; brand-voice policy added to evals; no AdoptionState demotion required.

When Noah and Ines became Customer Outcome Stewards, the disputes weren't a back-office cost any more — they were an outcome lever. That changed who they were in the org, in a good way.

Priya Raman, Chief People Officer
Audit reference: dr-bill-2025-q4
Rung 4 · Q2 2026 (current quarter)
20 weeks build · just released
AutonomousShipped

End-to-end agent with attestation cadence; humans on the loop

The agent owns the entire billing-to-collection cycle. Tom Whitfield stopped reviewing invoices and disputes individually — he now signs a quarterly attestation that the agent's pattern of behaviour is within policy, supported by the platform's continuous evidence pack. Auto-fallback to supervised mode kicks in if drift trips a threshold.

Cycle time
1.5d
-92%
Error rate
0.4%
-3.8 pp
Team size
3
-3
Monthly toil
£3k
−£37k/mo saved
Trigger
Objective
Cash conversion uplift — reduce DSO from 41 to under 28 days
CFO-level. The platform's projections showed an autonomous agent could compress DSO faster than any process change.
Scenario chosen
End-to-end agent + attestation cadence (autonomous)
Agent runs the full process. Tom attests quarterly. Continuous drift monitoring with auto-fallback. Customer-facing language reviewed in batches by Outcome Stewards.
£410k·4.7 mo·med
Alternatives considered
  • Stay at supervised-autonomous, tune furtherDiminishing returns. The supervision overhead was costing more than the residual risk.
Workforce impact

Tom Whitfield transitioned from Head of Billing to Agent Steward (Billing). His direct reports dropped from 4 to 2 (the two Outcome Stewards). The Billing OrgUnit was flagged as a candidate for terraforming in the next planning cycle.

Pathways
  • Tom Whitfield · Head of BillingAgent Steward (Billing) · completed
Emergent roles
  • Agent Steward (Billing) — approved, filled
Assessment
End-to-end billing agent — release gate
pass
brief conformance: passcorrectness: passoutcome: passworkforce-intent: passregulatory: passsafety: pass

Continuous evidence pack; quarterly re-attestation. First attestation due 2026-Q3.

I went from running a team that processes invoices to running an agent that processes invoices. Same outcome ownership, much less middle-management. I'd rather be a Steward.

Tom Whitfield, Agent Steward (Billing)
Audit reference: dr-bill-2026-q1
Rung 5 · Q4 2026 (planned)
Decision pending · planning in flight
TerraformedPlanning · decision pending

Continuous reconciliation — invoicing dissolves into real-time settlement

The Function 'Generate customer invoice' stops being a thing. Instead of monthly invoices, Helios moves customers to a real-time settlement model: every shipment settles against contract terms as it completes. 'Billing' as an OrgUnit dissolves; the work merges into Treasury (cash flow) and Customer Success (commercial relationship). The agent doesn't 'invoice' — it maintains a continuously reconciled position.

Cycle time
real-time
-100%
Error rate
0.05%
-4.2 pp
Team size
0
-6
Monthly toil
£0
−£40k/mo saved
Trigger
Objective
Reshape Helios's commercial model · 'invoicing as a thing' is the constraint
Recognition that the autonomous agent had made monthly invoicing the slowest part of cash conversion. Customers were ready for real-time settlement; competitors were starting to offer it.
Scenario chosen
Continuous reconciliation (terraformed)
Replace 'Generate customer invoice' with 'Maintain real-time settlement'. Renegotiate top-30 customer contracts. Dissolve Billing OrgUnit; merge into Treasury + CS. Workforce impact significant — pathways for all remaining specialists.
£1.35m·9.3 mo·low
Alternatives considered
  • Stay at autonomous, refine furtherHits the ceiling. The Function as conceived can't get faster than 'real time' which is what we'd be approaching anyway.
  • Roll out continuous reconciliation only to top 5 customers, keep monthly for everyone elseConsidered. May be the staged delivery path within the chosen scenario. Decision deferred to plan kickoff.
Workforce impact

Billing OrgUnit (current headcount 3) dissolves over 9 months. Tom remains Agent Steward but for the new continuous-reconciliation function. Outcome Stewards merge into Customer Success. 1 specialist will need a substantive transition pathway — current candidate destinations: Treasury Analyst (emergent) or CS Operations.

Pathways
  • Tom Whitfield · Agent Steward (Billing)Agent Steward (Settlement) · draft
  • Customer Outcome Stewards · Billing-attachedCustomer Success · draft
  • 1 remaining Specialist · Billing SpecialistTreasury Analyst (emergent) · draft
Emergent roles
  • Treasury Analyst (settlement-aware) — predicted, not yet proposed
  • Settlement Steward — predicted
Assessment
Continuous reconciliation — pre-decision diligence
marginal
brief conformance: passcorrectness: marginaloutcome: passworkforce-intent: marginalregulatory: marginal

Marginal because: contract renegotiation is the binding constraint, not the technology. Workforce-intent marginal because the dissolution of Billing OrgUnit is large change for 3 people. Regulatory marginal because real-time settlement crosses revenue-recognition lines that need finance/audit sign-off.

Terraform isn't 'more autonomy.' It's the moment you realise the question you were optimising was the wrong question. We don't need a faster invoice — we need to stop having invoices.

Margaret Holloway, CEO
Audit reference: dr-bill-2026-q4 (pending)
End-to-end recap

From 6 specialists & 18-day cycles to a continuously-reconciled customer position

Six rungs. Twenty-one months. Five signed decision records. Two emergent roles created and filled. Two drift events caught and resolved. The Function being optimised stopped existing — replaced with a new one that's native to autonomous agents.

Decisions signed
5
Emergent roles
2
Drift events
2
Total invested
£2.26m