Agentic Systems

Agentic systems for professional services, sales and gym operations

An agentic system uses an AI model, business context and permitted tools to carry out a bounded piece of work. The useful question is not whether it can hold a conversation. It is whether it can help someone finish a job, with enough evidence and control to trust the result.

Ironstone's proposed engagement starts with one consequential workflow. We map the work, decide what an agent may do, test it against real examples and make a decision about production. Frontier models create new possibilities. The architecture still has to account for customer value, data, cost, security and the people responsible when something goes wrong.

Professional services

Start where skilled people spend time reconstructing context: understanding an enquiry, preparing a scope, finding relevant project material or checking a draft deliverable. An agent could assemble the evidence and prepare the next step. The professional remains responsible for the advice, the client commitment and the final work.

A useful pilot might prepare a project briefing from an approved collection of documents, cite its sources and flag what it cannot establish. It should make the practitioner better informed, not merely produce more pages.

Sales

Research, meeting preparation and follow-up are connected activities. Too often the knowledge gathered for one conversation disappears before the next. An agent can help maintain that context, distinguish a confirmed fact from a sales hypothesis and prepare work for review.

A first pilot could turn approved account material and meeting notes into a sourced briefing, proposed next actions and draft CRM updates. External messages, pricing and contractual commitments stay behind explicit approval. More activity is not the objective. A better-informed customer conversation is.

Gym operations

A gym has a practical service business around the training: enquiries, memberships, schedules, class changes and follow-up. An agent could help staff answer routine questions from approved information, prepare a booking or identify an unresolved service request. Escalation to a person needs to be easy.

The starting scope is business administration and member service. It excludes health information, medical advice and automated decisions about a person's fitness. Before connecting tools, we decide which records are needed, who may see them and which changes require approval.

What a first engagement would produce

1. A workflow and value assessment. Identify the user, the job, the current baseline, the cost of failure and the decision worth improving.

2. An architecture and authority model. Define approved data, tools, identities, permissions, human approvals and the system's stopping conditions.

3. A bounded pilot and evaluation. Test representative work, exceptions and failure cases. Examine usefulness, factual support, latency, cost and the effort of human review.

4. A production decision. Decide whether to proceed, revise or stop. If proceeding, agree ownership, monitoring, change control, fallback and support before expanding access.

The pilot can also show that an agent is unnecessary. A clearer process, a tested calculation or a simpler integration may be the better answer.

How we choose the technology

Start with the work and evaluate suitable current frontier models against it. Keep model selection separate from the business rules, permissions and evidence that the organization owns. A change of model should trigger evaluation, not a leap of faith.

Ironstone's related work includes NeoCRM, an active experiment and reference architecture, and NeoLibrary, a method for turning accumulated material into working knowledge. They inform the thinking. They are not presented here as finished commercial products.

Bring one workflow

Show us the work that matters, who does it and what a good result looks like. We can start there, before anyone commits to an agent platform.

Book a 30-minute conversation Email Rob