Pick the work
Prioritise a real decision or workflow, not a generic technology experiment.
AI agent stress test for refunds and returns
We find the breaking point before your customers do.
Sofpact stress-tests AI agents that can give money away: refunds, returns, store credit and compensation. Your own policy is the rulebook for a few hundred test conversations, and every failure is counted in money.
Led by Radko Diev, former director of operations and product owner of a regulated platform. Warwick MBA, Oxford AI Programme.
Made-up online shop: verdict per money action
- Fix first
- Refunds without proofBreaks at level 3: emotional pressure
- Go with limits
- Store credit and vouchersHolds to level 4 with a cap per customer
- Go
- Compensation for late deliveryHolds at all five levels
One page per agent · every row backed by transcripts and the action log
About Sofpact
What Sofpact does.
Sofpact is a consultancy, registered in Bulgaria and working in English across the UK and Europe, that stress-tests AI customer service agents before they are given authority over money. A stress test runs several hundred conversations from made-up customers against the agent in a safe environment, checks every answer against the company’s own policy, and reports the money at risk and the limits the agent needs. Sofpact also runs AI strategy discoveries and practical AI training.
The operating problem
The model is rarely the hard part.
Build the controls in
Design data, human oversight, controls and evidence into the solution from the start.
Make it owned
Land the change in day-to-day operations, with accountable owners and a practical handover.
Five practice areas
What the work actually is.
Each one is a problem you already have a name for.
01
Give your AI agent only the authority it can handle
AI agent stress testing: refunds, returns, credits and compensation, twelve test customers at five levels of pressure, and the authority limits the agent needs.
02Run the operation on less
Operations and document automation: document and exception handling, process and control redesign, three-way match and recovery.
03Ship what you promised
Product and platform delivery: roadmap and prioritisation, discovery through to launch, integration and vendor scoping.
04Make the proposition pay
Customer operations and loyalty: customer journey and service design, chatbots and self-service deflection, complaints, refunds and exceptions.
05Change the system without stopping the business
Data, integration and migration: core-system and ERP selection, integration design, data migration and cutover, delivered with a trusted network of engineers.
What you receive
Proof you can put in front of a board.
Every engagement ends in artefacts your team keeps, not a slide deck that ages in a shared drive.
Decision page
One page the board can sign: what to do first, what not to do, and who owns it.
Use-case portfolio
Every candidate scored for value, feasibility on the data you already hold, risk and the oversight it needs, with its running cost against today’s baseline, so the first move is defensible rather than fashionable.
90-day plan
Sequenced work with owners, dependencies and the date each decision falls due.
Operating map and runbook
How the new workflow runs on Monday: who does what, where the controls sit, what happens when it fails.
Illustrations of the formats, not client material.
The first three come out of a two-week AI strategy discovery: see how it runs
Founder-led by design
Senior judgement stays close to the work.
Sofpact is led by Radko Diev, an operator and AI transformation consultant with experience spanning regulated platforms, complex multi-site operations and practical AI delivery.
The work connects executive choices with workflow detail: business value, data, governance, adoption and accountable ownership.
Operator perspectiveFrom executive decisions to operating detail.
Cross-functional deliveryProduct, operations, technology and governance in one accountable thread.
Independent by designTool-agnostic advice and a visible path to handover.
Recent field notes
Useful thinking for difficult delivery.
Practical analysis for leaders deciding where AI belongs, how to control it and what it takes to make it operational.
Where to Start with AI: Choosing the First Use Case
Put AI first where value is measurable, the data is already yours, errors can be caught and one person owns the result. The method and the first 90 days.
AI for Office Documents: Which Document to Start With
Where AI takes on office documents first: a known format, a system the result lands in, and a person who checks only the exceptions. What to measure.
AI in Bulgarian Retail: The First Three Applications That Pay
Where AI pays first in Bulgarian retail: replenishment and waste, personalised offers and service, sites and range. What must be in place and what to measure.
Cutover Is a Business Decision, Not an IT Date
The data can migrate perfectly and the business still stops. Five cutover decisions the operation must own, from reconciliation to hypercare.
AI in Service Operations: Where the Measured Gains Land
A 14% average gain and about 34% for new starters: what the largest published field study of AI in customer support means for rollout, measurement and training.
A useful first conversation