SAP to open-stack ERP migration
Manufacturer, 500+ employees. Phased migration off SAP ERP onto an open stack with custom modules — run in stages while the plant kept working.
- 0 downtime
- 8 mo migration
- −40% on licences
AI agents that do the work — with an audit trail
An agent is not a chatbot with ambition. It reads your systems, takes multi-step actions, asks for approval where the stakes are real, and leaves a record of every decision it made.
What we can show you
Overview
Agent or chatbot: the line that matters
An agent does work rather than talk about it — which is exactly why it needs approval gates, a full trace and a kill switch.
What separates an AI agent from a chatbot
An agent does work rather than talk about it: it reads your systems, calls tools in sequence, decides what to do next, and stops for approval where the stakes are real. The conversation, if there is one, is a side effect.
That autonomy is exactly why an agent needs controls a chatbot does not: typed tool calls, approval gates, a full trace of every step, budgets enforced in code and a kill switch that works without a deploy.
When an agent is justified
When a process is repetitive, rule-bounded and high volume, and when each step can be reversed or approved. If the work needs judgement that nobody has written down, the first project is writing it down — not building an agent.
Signals a process is ready for an agent
- someone senior spends hours a day on a routine they can describe step by step;
- the inputs arrive as documents, tickets or records rather than conversations;
- the systems involved have APIs, or can get a thin one;
- mistakes are reversible, or a human can approve before the irreversible step;
- volume is high enough that a 30% deflection is worth a project.
Price
From $18,900
Timeline
4–6 weeks
Payment
50/50 by milestone
Comparison
An agent, a script, or a person
Three ways to handle a repetitive process, and the honest trade-off between them.
Criterion
Script / RPAAI agent
AI agentDeterministic script or RPA
Handles unstructured input
Breaks on anything unexpected
Reads documents, tickets and free text
Predictability
Total — same input, same output
High with evals and gates, never absolute
Maintenance when the source changes
Rewrite the rules
Update prompts and rerun the eval set
Auditability
Code is the audit trail
Full trace per step, stored and searchable
Cost per run
Near zero
Tokens, capped in code
When it wins
Fixed formats, strict rules, high volume
Messy inputs, judgement inside written rules
Scope
What we build
Four families of agent, all of them measured on completed tasks rather than plausible conversations.
Back-office agents
Document intake, matching, exception triage, data clean-up. The work that is high volume, rule-bounded and currently done by someone senior enough to resent it.
Sales and CRM agents
Enrichment, qualification, follow-up drafting — with a human pressing send. The draft is the product; the sending stays yours.
Operations agents
Monitoring, dispatch suggestions, inventory and reorder proposals, with the reasoning attached so a dispatcher can disagree.
Internal developer agents
Test generation, migration scaffolding, review assistance — inside your pipeline, reviewed like any other change.
Architecture
The parts we insist on
Six controls that separate an agent you can leave running from one that needs babysitting.
Non-negotiable in every agent we ship
Tools, not screen-scraping. Explicit, typed tool calls against your APIs — never a bot pretending to be a browser user.
Approval gates. Anything irreversible — money, customer communication, deletion — waits for a human unless you explicitly opt out per action type.
Full trace. Every step, tool call, input and output stored and searchable. When someone asks "why did it do that?", there is an answer.
Budgets in code. Token, latency and action-count limits enforced per run.
Evaluation in CI. Multi-step tasks scored end to end on a labelled set, with regression gates before merge.
A kill switch. One toggle that stops the agent everywhere, plus a documented rollback.
Single-purpose agent
From $18,900
Timeline
4–6 weeks
Feasibility review
$4,490, one week
Why us
Why teams choose us
Four things we can prove, not four adjectives.
Approval gates by default
Money, customer communication and deletion wait for a human unless you explicitly opt out per action type. Autonomy is a setting, not an ideology.
Every step traceable
Inputs, tool calls and outputs are stored and searchable, so "why did it do that?" has an answer rather than a shrug.
Budgets and a kill switch
Token, latency and action limits enforced in code, plus one toggle that stops the agent everywhere and a documented rollback.
Evaluated end to end
Multi-step tasks scored on a labelled set with regression gates — a change that breaks step seven does not reach production.
Industries
What we buildfor your industry
Agents earn their keep in document-heavy operations: logistics, finance back office, insurance and professional services.
- IT companies and digital agencies
- Insurers and brokers
- Lenders and microfinance
- E-commerce
- Wholesale and distribution
- Manufacturing
- Freight and forwarding
- Clinics and medical centres
- Dental
- Car dealers and service centres
- EdTech
- Legal and consulting firms
- Equipment rental
- Real estate agencies
- Property developers
- Travel and events
- Beauty and wellness
- Construction and renovation
FAQ
Agent questions
What comes up once the conversation moves past the demo.
How much does AI agent development cost?
From $18,900 for a single-purpose agent; multi-step workflows with approvals start at $35,900. A one-week feasibility review at $4,490 de-risks the estimate and is credited against the build.
What can an agent actually be trusted with?
Reversible, high-volume, rule-bounded work. Irreversible actions sit behind approval gates by default — that is a design decision, not a limitation we are apologising for.
How do you stop it doing something expensive?
Budgets and action limits enforced in code, approval gates on the risky classes, and a kill switch that works without a deploy.
How do you evaluate something with ten steps?
End-to-end task scoring on a labelled set, plus per-step traces so a failure can be localised to the step that caused it.
Which frameworks do you use?
Whatever survives your constraints. The architecture keeps orchestration and model swappable, and we do not build on a framework we cannot debug at senior level.
Can it work with our legacy systems?
Usually via a thin API layer we build first. If there is no API at all, that layer is the actual project — and we will say so before quoting the agent.
2–4×
faster on the code-writing share
in 3 weeks
from idea to MVP
230+
projects delivered
17 years
of engineering experience
Work
Work we can show you
SMARTWILL
Digital inheritance service (Web3)
AIRDEVELOPPA
Indoor air quality finder driven by sensor data
CAR5
Car-sharing app
APP.DENTAL
Cloud CRM for dental clinics
NEXORA
Warehouse and logistics management platform
As-built docs
Digital as-built documentation
EX211
Telegram Mini App for currency exchange
SMARTWILL
Digital inheritance service (Web3)
Bring the task, not the buzzword
Describe one repetitive process with a clear definition of done. We will tell you whether it is an agent, a script, or a rule that never needed a model.
Feasibility review is one week and credited in full against the build.