Intelligent automation,
AI in the loop.
Blends of deterministic automation with LLM reasoning, where the boring 80% stays boring and reliable, and reasoning is applied to the 20% that genuinely benefits from it.
The intelligence in intelligent automation is knowing where reasoning helps and where it hurts. The healthy programmes engineer both with equal seriousness, and are honest about the difference.
Where reasoning helps
Classification, extraction, summarisation, triage, drafting, where variance and context defeat rules.
We identify the steps where a stable rule cannot cover the variance, and where a reasoning model has evidence it will do better. That's the honest scope of AI in automation.
Everywhere else we keep deterministic logic, because it is cheaper, faster and more reliable.
Adding an LLM to a stable rule-based step usually just makes it slower, more expensive and less reliable.
Design the seam
Deterministic flow around reasoning steps. Structured input, structured output, human-in-the-loop where blast radius matters.
Reasoning steps live inside deterministic orchestration. Inputs and outputs are typed. Confidence thresholds route to human review where risk warrants.
Failures are diagnosed, not guessed, because every step is traceable.
Evaluate and operate
Task-level and step-level evals in CI. Quality, cost and drift monitored per surface.
Eval sets built from real cases. Regressions caught before deploy. Human review outcomes flow back into evals.
The systems that age gracefully are the ones evaluated continuously.
Scale
Portfolio economics measured monthly. Reasoning steps that stop paying back get replaced with rules.
Value tracking distinguishes automations that pay back from ones that just moved the work. Reasoning steps that could be replaced by rules, because the variance dropped, get replaced.
That's how the portfolio's economics stay defensible over time.
Toolkit
Framework, platform and vendor selection is driven by the shape of the work and the team who'll own it, not by whichever ecosystem is loudest this year. The list below is representative, not prescriptive.
- Orchestration
- Camunda · Temporal · Airflow
- Reasoning
- Anthropic · OpenAI · LangChain
- IDP
- ABBYY · AWS Textract · Azure Doc Intel
- Evaluation
- LangSmith · Langfuse
- Platform
- Kubernetes · AWS · Azure · Datadog
Selected work
What clients ask us first
01Where does reasoning actually help?
Classification, extraction, summarisation, triage and drafting, where variance and context defeat static rules. Everywhere else, deterministic logic is cheaper and more reliable.
02How do you design the seam between deterministic and reasoning?
Reasoning steps live inside deterministic orchestration. Typed inputs and outputs, confidence thresholds, human-in-the-loop where risk warrants. Failures are traceable.
03How do you evaluate quality?
Task-level and step-level eval sets from real cases. In CI, gating deploys. Human review outcomes flow back into evals so the set keeps up with reality.
04How do you keep economics healthy?
Live portfolio value tracking. Reasoning steps that could be replaced by rules, because variance dropped or the domain stabilised, get replaced.
Let's discuss your intelligent-automation programme.
Share a brief overview of the processes you're automating, we'll arrange an introductory conversation with the relevant practice leadership.