88%
routine-query deflect rate at a federal health body after one quarter in production
4 days → 1
wait time reduction at a federal health engagement — 8,000 queries a month
23%
response error rate found at Diagnose — caused by stale templates, not poor staff performance
240 hrs
staff hours freed per week within one quarter of the reply helper going live
The core problem
Your support team isn't slow. They're working from templates last updated before the last policy change.
A federal health body ran 8,000 staff queries a month. Four-day wait times. Headcount at capacity. Diagnose read 12 months of tickets and found that 23% of outgoing responses contained an error traceable to a policy change that was applied to the source guidance but never reflected in the reply templates. The team was not performing poorly — they were working from stale material at volume. That is a data problem, not a staffing problem.
The structural fix is the same across any support-heavy operation: AI reads the incoming query, drafts a response trained on your current policy and past resolutions, and presents it for a human to review and send. When policy changes, the training update propagates to every future draft rather than relying on each agent to notice the memo.
What changes
The same challenge. Two very different outcomes.
Without Effektiv
- 4-day wait times with headcount at capacity
- 23% error rate in outgoing responses from stale templates
- Staff time consumed drafting from memory rather than current guidance
- Policy changes propagated manually, inconsistently, slowly
- No visibility into which queries drive the most error risk
- Human sent every response — with no draft assistance
With Effektiv
- 1-day wait time after one quarter in production
- Under 2% error rate — AI surfaces current guidance with each draft
- Staff focus on review and judgment, not drafting from scratch
- Policy changes propagate to every future draft automatically
- Escalation precision dashboard — live visibility for supervisors
- Human in the loop on every send — a contract clause, not a setting
How we deliver
Diagnose. Design. Deliver.
Two weeks of listening before a line of code. The price is fixed at the end of Design — not at kick-off.
Phase 1 · 1–2 weeks
Diagnose
We read 12 months of ticket history, identify the most common query categories, and audit reply accuracy against current guidance. This is where stale-template errors are found — before they compound further. The error rate and category breakdown become the baseline for the outcome contract.
Phase 2 · 1–2 weeks
Design
Agent structure, human-gate rules, eval gates, and the weekly reporting schedule. Human-in-the-loop is required by design on any response that goes to a customer. All model inference on AWS Bedrock in AU regions. IRAP path available for PROTECTED data work from Q4 2026.
Phase 3 · 6–12 weeks
Deliver
The outcome contract names the deflect rate target, response accuracy target, and wait-time target — each measured weekly against the baseline set in Diagnose. The quality checks, reply templates, and prompt rules are yours at exit.
Quality gates
What the quality checks measures.
Every output passes a multi-gate evaluation before it merges or ships. Outputs that fail do not proceed. The quality checks and all gate code are yours at exit.
- Brand-voice match score — threshold agreed in Design against samples from your prior responses
- False-reply rate — target zero on customer-facing outputs. Drafts that fail go to escalation, not to send
- Policy match rate — scored against the live policy corpus, with citations attached to every draft
- Response-time delta vs the prior baseline, measured weekly against the contract target
- Escalation precision — correct escalations divided by total escalations. Drift triggers an eval refresh
Eval rig · sample run
Eval rig source code shipped to your repo at exit.
Sample engagement
A federal health body ran 8,000 staff queries a month with a four-day wait time and headcount at full capacity. Diagnose found a 23% response error rate driven by stale templates. The reply helper was built in six weeks, trained on current guidance with a human-in-the-loop gate on every send. Within one quarter, 240 staff hours a week were freed, wait time fell to one day, and the routine-query deflect rate reached 88%. IRAP sign-off in progress for PROTECTED data work.
Read the full case →
Compliance posture
ISO 27001 in progress (Q3 2026) ISO 42001 aligned NIST AI RMF mapped IRAP path Q4 2026 Full governance posture →Other services
Other ways we work with you.
Service
Modernisation
AI archaeology decodes what the documentation missed. 11 weeks median.
Read more →Service
Operations & Integration
RPA and ESB out, connected AI systems in. Run cost down 50–70%.
Read more →Service
AI Adoption
With-you mode. Your team ships AI without us in 90 days.
Read more →Service
Software Build
Custom products and greenfield software, shipped in 8–14 weeks.
Read more →Service
Maintenance & SRE
AI in the alert path. 50–70% drop in human-paged incidents.
Read more →Common questions
Frequently asked questions.
Human in the loop. Always.
See what your support operation looks like with AI doing the drafting.
Show us your ticket volume, your current wait times, and your policy corpus. We price on outcomes: deflect rate, accuracy, and wait time — all measured weekly against the baseline we find in Diagnose.