The most common objection we hear when a business first considers AI agents is not about cost, and it is not about trust. It is this: "We just finished implementing our ERP two years ago. We are not doing that again."
Fair. ERP and CRM implementations are painful, expensive and disruptive, and the idea of another migration is enough to end the conversation. The good news is that in most cases it is also unnecessary. Agents do not require a new system. They require access to the one you have.
The Principle: Agents Are a Layer, Not a Replacement
An AI agent needs three things to do useful work: a goal, tools it can operate, and permission to act within limits. Your ERP or CRM is the tool. If it exposes an API, webhooks, or a database we can connect to safely, an agent can read from it, act through it, and write results back to it — exactly as an authorised member of staff would, but continuously.
Your team keeps the screens they know. The records stay in the same place. Nothing gets migrated. The difference is that the work now happens without someone doing it by hand.
Systems That Carry Agents Well
- ERP — Odoo, SAP, Microsoft Dynamics, Sage, and most custom in-house systems
- CRM — HubSpot, Zoho, Salesforce, Pipedrive, Freshsales
- Finance — QuickBooks, Xero, Paystack, Flutterwave, Stripe
- Operations — Google Workspace, Microsoft 365, Slack, Notion, Airtable
- Channels — WhatsApp Business API, voice, email, web chat, SMS
If your system is not on that list, the question is simply whether it has an API, webhooks, a database we can safely connect to, or even a structured export. There is usually a route. Occasionally there is not, and we will tell you that plainly rather than sell you an integration that will be fragile forever.
How the Build Actually Goes
Step 1 — Process walkthrough, including the workarounds
We map the process as it genuinely works today, not as the manual describes it. The informal workarounds matter most: the spreadsheet someone maintains on the side, the WhatsApp group where the real approvals happen, the step everyone skips. Agents deployed onto the documented process rather than the real one fail immediately.
Step 2 — Define the agent's job, authority and escalation
Before any code: what is this agent's goal, which tools may it use, which records may it touch, what may it do alone, what requires human approval, and what does it do when it is unsure. This document is the actual product. Everything else is implementation.
Step 3 — Integration and business memory
We connect the agent to your systems with scoped credentials — read and write access only to what the job requires — and load its business memory: your policies, pricing rules, product catalogue, service scope and relevant history. This is what makes it answer as your company rather than as a generic model improvising.
Step 4 — Test against real history
We run the agent against your actual past cases, including the awkward ones. Not synthetic examples. Real enquiries, real exceptions, real messy inputs — and we compare what it would have done against what actually happened.
Step 5 — Supervised deployment
The agent goes live on real traffic with a human reviewing its actions before they take effect. You watch how it behaves on your business, in your market, with your customers, and adjust before it runs unsupervised. This phase typically lasts one to three weeks and it is not optional.
Step 6 — Expand
Once one agent is proven against real numbers, the next is easier and faster: the integrations exist, the business memory exists, and the internal trust exists. Most clients go from one agent to a departmental agentic layer within two quarters.
What If Your System Cannot Carry Agents?
Some systems genuinely cannot: no API, a closed vendor platform, or a data model so tangled that safe integration is not realistic. Others can technically carry agents but should not, because the business has outgrown the system anyway and you would be investing in something you are about to replace.
In those cases we build the agentic system itself — a custom ERP, CRM or operations platform designed around agents from day one. That means proper event architecture, clean data models, tool interfaces agents can operate, and permission scoping and audit logging built in rather than added later. It is a bigger project, and we only recommend it when the alternative genuinely does not work.
What We Need From You
Less than most clients expect: access to the relevant systems, an honest walkthrough of how the process really runs, your business rules and policies, and one internal owner who can answer questions and approve agent behaviour during the supervised phase. We handle design, engineering, integration, testing and rollout.
How Long and How Much
A single agentic workflow deployed into an existing system typically goes live in 2 to 4 weeks and starts from ₦850,000 for design, build, integration, testing and supervised rollout, plus a monthly running cost for model usage, hosting and monitoring. A multi-process agentic layer across a department typically takes 8 to 14 weeks and starts from ₦3m depending on integration complexity and data quality.
The comparison worth making is not against software pricing. It is against what the manual version of that process currently costs you in salaries, delay and error.
Start With One
Do not attempt an agentic transformation. Pick the single process that is repetitive, high-volume, time-sensitive and currently eating people who should be doing something else. Deploy one agent there, supervised. Measure it for six weeks.
That is a small, reversible, honest experiment — and it settles the question with your own numbers instead of anyone's marketing.
Harzotech builds agentic workflow systems inside existing ERP and CRM platforms across Nigeria and Africa. Tell us the process that eats your team's week and we will tell you honestly whether it can be made agentic on the systems you already own.