AI Agents · Scoring · Routing · Summarization

AI agents for clearly defined business tasks.

Sort incoming requests, score leads, and summarize calls with agents connected to your existing tools. We define the rules, test against examples, and add human review where decisions need it.

✓ Scoring · routing · summarization✓ Frontier LLMs · structured outputs✓ Confidence scores · human gates✓ Wired into CRM · Slack · email

Illustrative example. The final product is designed around your requirements.

Best fit

Businesses building a product or improving their tools.

  • New products that need design, development, and launch support.
  • Internal tools and CRMs that bring scattered work together.
  • Founders who need a team to plan and build with them.
How we work

A shared plan from brief to handover.

  • Agree the requirements, scope, and budget before the build.
  • Review the work together throughout development.
  • Document the product and the tools your team will maintain.
On the first call

Work out what to build first.

  • What to build first and what can wait.
  • Which tools and development approach suit your project.
  • An initial view of scope, timing, and budget.
What we build

Agents built around specific tasks.

These are the agents we build most often. Each one replaces a chunk of human judgment that doesn't need to be human.

01

Lead scoring agents

Reads the form / chat / call transcript, scores intent + fit against your ICP, ranks 0–100 with reasoning. Hot leads flagged to AEs in Slack within 30 seconds.

02

Routing & triage agents

Incoming tickets / emails / DMs classified by topic, urgency, and language, then routed to the right human (or auto-resolved if simple). Response time drops 80%.

03

Summarization agents

Sales calls, support chats, internal meetings - summarized with action items, sentiment, and follow-up drafts. Everything stored in CRM with the transcript attached.

04

Decisioning agents

Approve / reject with reasoning: refund requests under $X, discount asks, content moderation, access grants. Low-stakes decisions auto-execute; high-stakes escalate with a recommended action.

How we work

Test the decisions before automating them.

We check results against past examples and review live decisions before allowing the system to act.

01

Define the task

Week 1. Document the information needed, decision rules, exceptions, and human review points.

02

Prepare examples

Weeks 1–2. Collect 100–500 past examples with the correct answers to use in testing.

03

Build and test

Week 2. Build the agent, connect its tools, and check each change against the examples.

04

Compare with your team

Week 3. The agent records its decisions without taking action. We compare them with your team’s decisions.

05

Introduce it gradually

Week 4. Start with clear cases and send uncertain ones to a person. Expand from 20% to 50% to 100% as results meet the agreed checks.

06

Monitor and update

Keep a record of decisions, review samples, and update the system quarterly.

Tools

Tools we use.

Models
GPT-5Claude OpusClaude HaikuGeminiFine-tuned open-source
Orchestration
LangGraphAnthropic SDKOpenAI SDKFunction calling
Evals & observability
BraintrustLangSmithCustom eval setsSentryLogfire
Data · memory
PostgresPineconepgvectorSupabaseEmbeddings
Integration
HubSpotSalesforceSlackGmailIntercomZendesk
Pricing

Per agent. No per-decision tax.

We charge to build and evaluate the agent well. LLM API calls are pass-through (usually $0.005–0.05 per decision). No markup.

Single agent · one decision type from $12k
Agent suite · 2–4 related decisions from $25k
Full decisioning layer · team-wide from $55k
Monthly retainer · evals + new scenarios from $2.5k/mo
Every agent ships with a curated evaluation set (100+ labeled examples), confidence thresholds, human-approval gates where needed, and 30 days of monitored shadow + gradual rollout. LLM costs at pass-through.
Questions worth asking

The short version.

What if the agent makes a wrong call?

We engineer for safety. Three guards: (1) confidence thresholds - below X, escalate to a human; (2) structured outputs with validation - the agent can't return malformed data; (3) shadow mode for 2 weeks before any autonomous action. For high-stakes decisions (refunds, legal, medical), the agent recommends; a human approves.

How do you measure accuracy?

Every agent has an evaluation set - 100–500 labeled historical examples. On every change to the agent, we re-run the full eval set and compare. We don't ship changes that drop accuracy. You see the eval dashboard and can set the threshold we won't fall below.

Can it learn from new examples?

Yes - through two channels. (1) We add disagreements from shadow mode to the eval set, retuning the prompt. (2) For high-volume agents, we fine-tune a smaller open-source model on your labeled data, which often beats a frontier LLM on your specific task at 1/10th the cost.

What models do you use?

Depends on the task. GPT-5 or Claude Opus for nuanced reasoning; Claude Haiku or Gemini Flash for high-volume / cost-sensitive; fine-tuned open-source (Llama / Mistral) for domain-specific tasks where data privacy matters. We'll recommend per use case.

Can we keep data on-prem?

Yes. For regulated industries (health, finance, legal) we can deploy entirely on your VPC, using self-hosted open-source models via Ollama or vLLM. No data leaves your network. We've done it for clients with HIPAA, SOC 2, and ISO 27001 requirements.

How is this different from ChatGPT?

Agents do specific jobs with specific inputs, outputs, and accountability - not open-ended chat. Our scoring agent takes a lead object, returns a score + reasoning + confidence, logs the decision, and integrates with your CRM. It's a production service with SLAs, not a chatbot.

Ready to offload judgment calls?

Choose a task worth automating.

Book 30 minutes to discuss recurring decisions, the information they need, and where your team should stay involved.