AI agents that do the reading, and wait for a yes.
QServices builds AI agents that read emails, PDFs and ERP records, match them, and draft the next step inside the ERP and CRM you already run. Anything that posts money or changes a system waits for a named person on your team.
Built in your own Azure tenant, with Azure OpenAI or Claude models. Your ERP stays the system of record.
An AI agent is software that uses a language model to work toward a goal: it reads the situation, decides the next step and takes it by calling tools and systems, within the instructions and permissions you give it. A workflow follows steps someone wrote in advance. An agent chooses its own steps, which is why it can handle a messy email or an exception, and also why it needs limits, tests and a person on the decisions that matter. AI agent development is the work of designing, connecting, testing and running those agents.
Definitions follow Anthropic's "Building effective agents" (workflows run on predefined code paths; agents direct their own process and tool use) and Microsoft's Cloud Adoption Framework.Gartner calls it "agent washing": chatbots, RPA and assistants relabelled as agents. It estimates only about 130 of the thousands of vendors selling agentic AI are the real thing. Here's how to tell them apart.
| Generative AI (chat) | RPA bot | Workflow automation | AI agent | |
|---|---|---|---|---|
| How it decides | Answers what you ask | Replays recorded clicks and fixed rules | Follows steps you defined in a flow or code | Picks its next step itself, inside its instructions and permissions |
| Messy input | Reads it, doesn't act on it | Breaks when a screen or format changes | Only the formats you planned for | Emails, PDFs, scans and free text |
| Acts in your systems | Rarely | Yes, through the screen | Yes, through APIs | Yes, through the tools and APIs you allow |
| Predictability | Varies by prompt | High, until something changes | High | Lower, so it needs tests, limits and approval points |
| Best for | Writing, summaries, questions | Stable, high-volume screen work | Known steps in a fixed order | Exceptions, judgment calls, unstructured documents |
From IBM, UiPath, Anthropic and Microsoft Agent Framework documentation. Microsoft's own advice: if you can write a function to handle the task, do that instead of using an agent.
Microsoft's own Business Central Sales Order Agent works this way: it acts only within the role an admin assigns, creates a quote first because a quote doesn't touch planning or cash flow, and has a person approve every outgoing message. We build to the same pattern on any ERP.
We agree this with your controller before we build anything. It's the most useful page in any agent project, and the first thing an auditor will ask for.
Read-only, and only the tables its role allows.
Nothing changes in the books until a later step is approved.
Drafts sit in a safe state: an unposted journal, an unsent email, a quote that reserves nothing.
It moves money. COSO asks for human review scaled to the risk, with extra care on journal entries.
Your name is on it. Microsoft's Sales Order Agent has a person approve every outgoing message too.
The agent brings the history and a recommendation. Your credit manager decides.
A favourite target for fraud. The agent can flag a change request, never make one.
Judgment calls stay with the people accountable for them.
OWASP lists excessive agency among the top risks for AI applications. Only your admins change what an agent may do.
A starting point, agreed per process. Items can move left as trust grows, and the log shows when and why.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, for cost, unclear value or weak risk controls. MIT's research on enterprise AI pilots found most never reached the P&L. The reason shows up in where the work goes.
Cleaning the data the agent reads, connecting it to the ERP, agreeing the approval rules and fitting it into how the team works.
The part most demos show.
MIT Sloan (February 2026) reports 80% of the work on one healthcare agent went to data, stakeholder alignment, governance and workflow integration; the 20% is the remainder. It's why our projects start with your data and your approval rules, not with the model.
Five questions. Sometimes the honest answer is a plain workflow.
Does the work start from messy input: emails, PDFs, scans or free text?
Do the steps change from case to case, with exceptions that need judgment?
Does it happen a few hundred times a month or more?
Can a person check the result before it counts, as a draft or in a queue?
Does the data live in systems with an API or a connector?
Answer the questions to see what fits.
Messy input, judgment and enough volume: this is where agents pay for themselves. Start with one agent, one queue and one named approver.
Book a two-week assessmentIf the steps are known and the input is clean, a Power Automate flow or an integration is cheaper and more predictable than an agent.
Workflow automationIf nobody can check the work before it counts, or the data has no way in, an agent adds risk. Start with the integration or the approval step.
ERP integrationMicrosoft's own guidance is to check first whether an off-the-shelf agent, such as a Dynamics 365 agent, already does the job. When it doesn't, we pick one of these three.
| Copilot Studio | Microsoft Foundry Agent Service | Custom code | |
|---|---|---|---|
| Who builds it | Low-code. Your team can maintain parts of it | Pro-code, by developers | Developers, with the Agent Framework or the Claude and OpenAI SDKs |
| Good for | Agents in Teams, Outlook and Microsoft 365, with approval flows | Agents that call your ERP's APIs, with private networking and built-in evaluations | Unusual ERPs, heavy custom logic, agents across several clouds |
| Identity and data | Works within each user's Microsoft 365 permissions | Each agent gets its own Entra identity and runs in your Azure network | Whatever you build. You own every control |
| Models | Chosen and hosted by Microsoft | Azure OpenAI and Claude models, billed through Azure | Any model |
| How you pay to run it | Copilot Credits: $200 a month buys 25,000, or $0.01 each pay-as-you-go | No charge for prompt agents; model tokens and tools, plus container compute for hosted agents | Model tokens plus your own hosting |
| Watch for | Custom agents switch off at 125% of prepaid capacity | Needs Azure skills to run well | You maintain all of it |
Microsoft list prices and Microsoft Learn documentation, October 2026. Model prices change often; we quote yours after the assessment.
McKinsey's 2026 survey found 40% of companies with over $1 billion in revenue scaling agents, but only 22% of smaller ones, and no function above 12%. The ones that work share a pattern: the agent does the reading and drafting, a person makes the call.
| Department | What the agent does | What a person still decides |
|---|---|---|
| Finance | Matches payments to invoices, reads supplier invoices, drafts reminders and gathers dispute evidence. See cash application | Every posting, write-off and credit change |
| Sales and CRM | Turns an emailed order into an ERP quote and keeps CRM records clean. See sales agents | Pricing exceptions and what goes to the customer |
| Customer service | Sorts cases as they arrive and drafts "where's my order" answers from the ERP | Sends the reply and handles complaints |
| Operations | Reads supplier confirmations and flags late or short deliveries | Re-planning and the supplier call |
| IT operations | Triages tickets, watches integration syncs and writes incident notes | Any change to a production system |
| Healthcare admin | Assembles prior authorization packets and checks claims before they go out | Clinical judgment and every submission |
Healthcare: the AMA counts 39 prior authorizations per physician per week, about 13 hours of staff time. CMS requires payers to offer prior authorization APIs by 1 January 2027.
MIT's research found AI projects bought from specialist partners reached deployment about 67% of the time, roughly twice as often as internal builds (self-reported). These questions separate builders from demo shops.
Fixed price per phase. You see the agent working on your own documents before anything goes live.
We read your logs and documents, rank processes by hours and dollars, and pick the agent worth building first. See the process assessment.
Your controller or process owner agrees what the agent may do alone, what needs a person, and who that person is.
On Copilot Studio or Microsoft Foundry, with its own identity and least-privilege access, connected through your ERP's supported APIs. The ERP core stays untouched.
A test set of 20 to 50 real cases, including the ones that went wrong before. It reruns every time a prompt, tool or model changes.
The agent works real documents into a queue nobody acts on, until your team trusts the drafts.
Real work flows through the queue. Your approver says yes, edits or sends it back, and every decision is logged.
Our team runs the agent under one SLA, with a monthly report on volume, accuracy and overrides. The next agent reuses the identity, connectors and logging.
Running costs depend on volume and platform. For scale, Microsoft's own example puts its Business Central Payables Agent at about 6,500 Copilot Credits a month for 100 invoices.
Customers paid a seven-company services group by cheque, bank transfer, payment gateway and WhatsApp, including photos of payment slips. The AR team sorted every payment by hand.
The agent now reads each payment, works out the company and customer, matches the invoices and checks the bank log for duplicates. The AR team approves every posting.
Pick one that has shipped agents on your ERP, not just demos, and that will show you the approval matrix before building. The agent should run in your own tenant with its own identity, be tested on your real exceptions, and come with someone to run it after go-live. MIT's research found projects with specialist partners reached deployment about 67% of the time, roughly twice as often as internal builds.
There are two costs: building it and running it. We quote a fixed price per phase after the two-week assessment. For outside reference, published agency estimates start at about $25,000 to $75,000 for a single-task agent built in 4 to 8 weeks (DEV.co, October 2026). Running costs depend on volume: Copilot Studio sells 25,000 credits for $200 a month, and Microsoft's own example puts its Payables Agent at about 6,500 credits a month for 100 invoices.
Where the work starts from documents and ends in a posting a person can check: cash application, supplier invoice capture, collections reminders, dispute evidence and reconciliation prep. In each case the agent drafts and a person posts. COSO's 2026 guidance calls for extra oversight on anything touching journal entries and reconciliations.
Assembling prior authorization packets, checking eligibility, scrubbing claims before they go out, drafting denial appeals and answering billing questions. The AMA counts 39 prior authorizations per physician per week, about 13 hours of staff time. Clinical decisions and every submission stay with your staff.
Generative AI creates content from a prompt: a summary, an email, an answer. Agentic AI works toward a goal, and its output can be an action taken in another system, like a draft quote in your ERP. IBM puts it the same way. Most agents use generative models inside them.
RPA replays fixed steps on a screen and breaks when the screen or the input changes. An agent reads unstructured input, decides its next step and calls systems through APIs. RPA is still the cheaper choice for stable, rule-based work.
Only on low-risk steps like reading, matching and drafting. Anything that posts money, sends a message outside the company or changes master data goes to a named person first, and every step is logged.
Yes, which is why we design for it. In NIST's agent-hijacking tests, the strongest attacks succeeded 81% of the time, and the UK's NCSC says prompt injection may never be fully solved. So our agents get least-privilege access, treat document content as data rather than instructions, and can't move money without a person.
We deploy in your Azure tenant and region, using Azure-hosted models. Microsoft states that prompts and documents sent to Azure OpenAI in Foundry aren't shared with other customers or used to train its models. We use regional deployments, because Global and DataZone deployments can process prompts in other regions. Access to your systems is limited to what each agent's job needs.
SYSPRO, Epicor, Dynamics GP and Business Central, Infor, Sage and SAP Business One, plus Dynamics 365, Salesforce, HubSpot and Zoho on the CRM side. See the ERPs we support.
Start with one team, one agent and one named approver. Add the next once the first one has paid for itself.
Facts on this page checked on 22 October 2026.
Founder and CEO
Chief Sales Officer
Join our live webinar on Sept 24 and see five
ways to automate meetings, approvals, and
document search with tools already in your
tenant.
