What we do

You have the strategy. The operation has not changed.

Sofpact is led by an operator who was both director of operations and product owner of the same regulated platform, who has opened physical sites and shipped a consumer app, and who still builds the first working version himself. Five practice areas, one person accountable, and AI used only where it earns its place.

Choose the outcomes that sound like yours

AI agent stress testing

“We are about to let the agent approve refunds, and nobody can tell us what it does when a customer pushes back.”

  • Refund, return and credit limits
  • Twelve test customers, five levels of pressure
  • Money at risk per 1,000 conversations
  • Authority limits and where each one lives
  • A re-test after every model or policy change

Directed internal control at a multi-site business, and ran the support and anti-fraud work behind a loyalty app. Every verdict is Radko’s own.

How the stress test works

Add to your request Added to your request

Operations and document automation

“The same work piles up in the same places every month, and we only see what it cost afterwards.”

  • Document and exception handling
  • Process and control redesign
  • Three-way match and recovery
  • Anomaly detection across sites
  • Operating-model diagnosis

Directed internal control and risk-based audit across a multi-site operation with many departments, then diagnosed operating maps for enterprise clients as a transformation consultant.

Add to your request Added to your request

Product and platform delivery

“The roadmap keeps moving, and the integrations nobody scoped are what keeps slipping.”

  • Roadmap and prioritisation
  • Discovery through to launch
  • Integration and vendor scoping
  • ERP and core systems selection
  • UAT, release and cadence

Was both director of operations and product owner of the same regulated platform, from first specification to live service.

Add to your request Added to your request

Customer operations and loyalty

“We run a loyalty programme, the support queue never shrinks, and we cannot say whether either pays for itself.”

  • Customer journey and service design
  • Chatbots and self-service deflection
  • Complaints, refunds and exceptions
  • Loyalty design and economics
  • Coalition settlement and customer data

Ran the operating model behind a coalition loyalty app (the proposition, the settlement, and the support and anti-fraud operation underneath it) and took a second programme live with a group of retail partners.

Add to your request Added to your request

Data, integration and migration

“We have to move to a new core system, and nobody can tell us what breaks on the day we switch.”

  • Build, buy or partner decisions and ERP selection
  • Integration design and vendor scoping
  • Data migration and reconciliation
  • Cutover and parallel-run planning
  • Controls and audit trail built in

Ran a multi-country ERP rollout across several legal entities, and the provider integrations behind a regulated platform. Builds and migrations are delivered with a trusted network of engineers and data specialists, under the same accountability.

Add to your request Added to your request

Where AI actually sits

Not a sixth column, deliberately. The stress test checks AI agents themselves; in the other four areas AI is a technique, used only where it earns its place. The work is built so that most of it is ordinary, testable code that never touches a model, which is exactly what survives a review. Rules first, a model where the work needs reading or judgement, and an agent only where the work genuinely spans several tools. Every action is logged, and accuracy is watched after go-live rather than assumed.

Your request

Nothing added yet. Pick whichever sound like yours. More than one is normal, and you do not need to know the answer before you ask.

  • Give your AI agent only the authority it can handle
  • Run the operation on less
  • Ship what you promised
  • Make the proposition pay
  • Change the system without stopping the business
How far you want to take it

Not ready to send anything?Take the free AI diagnostic

What happens after you send it

  1. You send the situation

    One workflow, constraint or decision. Early-stage questions are welcome.

  2. A written reply within two working days

    Scope, duration and a fee. If it is not a fit, that is the reply, along with a view on who would be better placed.

  3. You decide what happens next

    Nothing is charged, and sending the request commits neither side. Any fee is quoted in writing and agreed separately, once the work has been scoped.

What it costs

Prices on request. Every quote is in writing, at a fixed price agreed before work starts. A working session is credited in full if an AI strategy discovery or Sprint is commissioned within 30 days. Sprints are quoted per phase once scoped; embedded support is billed as a fixed monthly fee. Clients in the EU are quoted in euros.

Where the work runs

On your own infrastructure by default. The validation logic is ordinary code and touches no model. Where a model is needed, your security team chooses local or EU-hosted.

Who does the work

Radko leads every engagement and stays in it. Builds, integrations and data migrations are delivered with a trusted network of engineers and data specialists. Their roles are agreed with you in the scope, they work under the same accountability, and nothing is sub-contracted out of sight.

How the work is bought

Four formats. Start at the smallest one that answers the question.

The five practice areas above are what you get. These are the shapes it comes in. Each one ends at a decision, and none of them assumes the next.

01

Working session

One day

A single working day on the problem, with a written summary within three working days. For when something is clearly wrong but not yet defined, or when you want to see how the thinking works before committing to more.

Fits whenThe problem is real but not yet named

02

AI strategy discovery

Two weeks

Interviews, evidence review and a decision workshop. Produces a prioritised set of use cases, each with an internal risk tier and the human oversight it needs, a readiness view across data, controls and ownership, the baseline each change has to beat, and a clear recommendation on what not to do. Already running pilots? The same two weeks decide which to scale, fix or stop.

Fits whenA board asking for an AI plan, several candidates or stalled pilots

How the two weeks run

03

Sprint

Two to six weeks

One workflow taken from mapped to working, tested with representative data, with human oversight and controls matched to its risk. Ends with the result measured against the baseline, what it will cost to run, and a scale, revise or stop recommendation you can put in front of a board.

Fits whenOne workflow is worth proving

04

Embedded support

Three months or more

Fractional or interim leadership where delivery crosses product, operations, risk and technology. Named outcomes, an owner in every affected team, training by role, a straight answer about what changes in each job, and an explicit handover.

Fits whenDelivery needs someone accountable inside

Testing an AI agent? The stress test has its own three sizes: a Check within a week of access, a full stress test in two to three weeks, and a re-test after each change. How the stress test works

Training a team? Training sits beside these four formats: a half-day leadership briefing and a one-day team workshop, built by role, on your own tasks and tools. How the training works

How the work runs

Evidence before expansion.

Every stage has a defined decision, visible artefacts and an exit. The aim is useful capability, not permanent consultancy dependency.

01 / Frame

Define the work

Outcome, workflow, users, constraints and accountable sponsor.

02 / Control

Design the guardrails

Data, risk, oversight, security and evidence requirements.

03 / Prove

Test in context

Working artefacts, user feedback and operational measures.

04 / Own

Embed and hand over

Roles, routines, documentation and the next decision.

FAQ

Common questions.

Who is Sofpact best suited to?

Operations, finance and product leaders in mid-market and scale-up organisations (accounting and legal practices, fintech and payments, retail and FMCG, multi-site operators) who need one workflow automated, one system changed or one decision made on evidence. Online retailers and consumer brands whose AI agent can approve refunds or credits start with the AI agent stress test.

Do we need a defined AI use case?

No. The AI strategy discovery is designed for organisations with several ideas, unclear priorities or open questions about data, controls and sign-off. A Sprint is more appropriate when one bounded workflow is already a credible candidate.

We already have AI pilots that stalled. Can you help?

Yes. An AI strategy discovery reviews the pilots in scope against the baseline each was meant to beat, finds what actually blocks production (data, integration, ownership, running cost or the model itself) and recommends scale, fix or stop for each. Stopping a pilot on evidence is a result, not a failure.

Does Sofpact build or only advise?

Both. Radko builds first versions himself. Production builds, integrations and data migrations are delivered with a trusted network of engineers and data specialists, with roles agreed with you in the scope, working under the same accountability. A Sprint ends with a working prototype and a scale, revise or stop recommendation.

Do you cover AI governance and risk?

Yes, as part of the work rather than as a separate policy document. Every AI strategy discovery, Sprint and embedded engagement names who owns each AI system and decision in its scope and sorts use cases by risk, so attention goes where the exposure is. Anything built or taken live also gets access matched to roles and data sensitivity, the human oversight and production monitoring it needs, including checks for errors, bias and drift, and an audit trail that holds the evidence an auditor asks for. Sofpact does not give legal or regulatory advice: your counsel and data protection officer own the legal interpretation, and Sofpact builds the controls that put it into practice.

Is Sofpact tied to a platform or model?

No. Recommendations follow the workflow, data, risk, integration and operating constraints. Sofpact does not require a preferred cloud, model or automation platform.

Where and in which languages do you work?

Sofpact works remotely across the UK, Bulgaria and wider Europe. The core website is available in English, German, French and Bulgarian. Engagements are normally delivered directly in English or Bulgarian; additional language support is agreed when required.

What does it cost?

Prices are on request. Every engagement is quoted in writing at a fixed price, set by its scope and agreed before work starts. A one-day working session runs remotely or on site and is invoiced after the day, and the fee is credited in full if an AI strategy discovery or Sprint is commissioned within 30 days. A two-week AI strategy discovery is priced by how many departments it covers and delivered remotely with expenses included. Sprints are quoted as a fixed price per phase once scoped, because the cost depends on the systems, integrations and data involved. Embedded support is billed as a fixed monthly fee, set by the days a week it needs, with a three-month minimum. Clients in the EU are quoted in euros.

Not sure which stage fits?

Start with the constraint, not the package.

Describe the situation