AI & Marketing · Consulting
Custom AI Agents & Assistants: Software That Finishes the Job, Not Just Chats About It
An agent earns its place when it completes a task end to end and hands a person something they can check.
The problem
What this usually looks like
A chatbot that answers questions is not the same thing as software that finishes a job. Your team still retypes line items out of supplier PDFs, still reads an eighty-page tender to find the four clauses that decide whether to bid, still copies lead detail out of email into the CRM by hand. Off-the-shelf assistants cannot help with any of it because they have never seen your price book, your terms or your last two hundred proposals. And when people do try them on real work, the failures are quiet ones. A confident wrong number in a quote costs more than no quote at all.
How I approach it
I build agents around one task with a defined start and a defined finish, rather than a general assistant that does everything adequately and nothing reliably. The agent answers from your own documents through retrieval, so every claim can be traced back to the page it came from instead of being recalled from training data. Anything that commits a price, sends to a customer or writes to a system of record passes through a human first. Before launch the agent runs against a set of real historical cases with known correct answers, and that same set is re-run after every model, prompt or document change so degradation shows up as a failed test rather than as a complaint.
Scope
What I actually do
Scope each agent to a single task with a clear finish line, quote drafted, tender summarised, lead entered, instead of an assistant nobody ends up trusting.
Build retrieval over your own material: price books, spec sheets, contracts, SOPs and past proposals, with the source cited on every answer.
Give the agent tools: CRM writes, calendar access, ERP lookups, email drafts, document generation, scoped to exactly what the task needs and nothing wider.
Put a human approval step in front of anything that sends, quotes a price or changes a customer record.
Build RFP and tender readers that pull scope, deadlines, bonding, insurance and compliance terms out of long documents into a one-page bid decision sheet.
Build inbox triage and CRM enrichment agents that classify, route and populate lead detail without anyone retyping it.
Build internal knowledge assistants over the company's own files so new staff stop interrupting senior people for the same twenty answers.
Stand up an evaluation set of real historical cases with known outcomes and re-run it after every change to catch silent drift.
Deliverables
What you get
- Agent specification defining the task boundary, tool access and escalation rules
- Deployed agent connected to your documents and existing systems
- Retrieval index over your own files with source citation on every answer
- Human approval queue for anything that sends, commits or writes
- Evaluation set of real cases with pass and fail scoring
- Activity log and review dashboard showing what the agent did and where a person overrode it
- Runbook and training for the staff who supervise it
Outcomes
What changes
- Repetitive document work handled without a person retyping it
- Answers traceable to a source document instead of asserted from nowhere
- Wrong answers caught at an approval step rather than by a customer
- Performance measured against a fixed test set, so quiet degradation is visible
- The same team absorbing more volume without adding headcount
Engagement
How we work together
Agent feasibility assessment on a single workflow
Build-and-deploy project for one agent
Phased rollout across departments
Ongoing evaluation and tuning retainer
Questions
Straight answers
How is this different from uploading our documents to ChatGPT?
An uploaded file is read once inside one conversation. An agent runs on a trigger or a schedule, retrieves from an index that stays current as your documents change, uses tools to write into your systems, and logs what it did so the work can be audited. The chat window is the interface, not the product.
What stops it from making things up?
Retrieval with citation, so an answer has to point at a source page; a narrow task boundary, so there is less room to improvise; and a human approval step on anything consequential. Where the agent cannot support an answer from a document, it is built to escalate rather than fill the gap.
How do we know it still works correctly in six months?
A fixed set of real cases with known correct answers gets re-run after every model, prompt or document change. Drift shows up as failed cases in a report you can read, which is the only reliable way to catch a system that degrades quietly rather than breaking loudly.
What kind of work is actually worth building an agent for?
High volume, repetitive, document-heavy and low in judgement. Reading tenders, drafting quotes off a price book, pulling fields out of packing slips, triaging inbound email. Work that turns on negotiation or relationship judgement is a poor candidate and I will say so.
Related
Services that usually go with this
AI Implementation & Automation Consulting
Self-hosted models, real workflows and measurable hours saved. No hype, no per-token surprise bills.
Business Workflow Automation
Find the work being done by hand, count the hours, automate it, then count the hours again.
Private & On-Premise AI Deployment
If the data cannot leave the building, the model has to come to it.
Hiring instead of contracting? See the full-time roles, or the ATS-formatted resume.
Contact
Tell me what you are hiring for
- Call: +1 519 278 5085
- Email: tim@timarmstrong.ca
- Based: London & Southwestern Ontario, Ontario, America/Toronto (Eastern Time)
- Availability: Available nationally across Canada, on-site, hybrid or remote
- Serving: London · Stratford · Woodstock · Kitchener: Waterloo · Cambridge · Guelph · Brantford · Sarnia · Chatham-Kent · Windsor · Toronto · Mississauga and Canada-wide