AI agents promise a great deal, but putting them into production raises hard questions. How do you control what the AI costs you? How do you know when to trust its decisions? And what happens when it gets something wrong?
At IT Integrated Business Solutions we have built AI systems for fraud detection, spam filtering and invoice processing, and taken them into production. One lesson keeps repeating: the most effective approach is not to send everything to the largest language model available. It is to build a workflow that uses the right technology for each task, checks the results, and brings in people when the stakes are high. The LLM should be the last step, not the first.
Done well, that means faster processing, lower running costs and automated decisions you can trust.
These figures come from systems we built, not from demos. Each was measured on one system under the conditions described below; what is achievable for you depends on your data and your workflow.
Have a workflow that is costly, slow or hard to trust? Tell us about it.
Discuss your AI project →The approach: a cascade, not a single model
An AI workflow that runs reliably in production usually looks less like “ask the model” and more like a series of filters, each more capable and more expensive than the one before it:
Deterministic checks
Schema validation, blacklists, lookups, arithmetic, business rules.
Catches simple cases and errors immediately, at almost no cost.
Microseconds · near-freeSmall, fast model
A classifier, embedding match or scoring engine that answers one narrow question.
Handles routine cases quickly and economically.
Sub-second · very cheapLarge model
An LLM for the cases the earlier stages could not decide confidently.
Expensive AI only where it is genuinely needed.
Seconds to minutes · costlyHuman review
For outcomes that remain uncertain, or where an error would be expensive.
Keeps people in charge of the decisions that carry real risk.
Slowest · most expensiveEach stage only sees what the previous one could not resolve. The aim is to resolve routine cases at the lowest practical cost, and spend more only where a decision is worth it.
Three systems we have built show how this works in practice.
Transit Ledger: invoice processing on hardware you already own
Finding invoices and receipts in a business’s email, extracting the details and filing them is tedious manual work, and the data shouldn’t leave the business. But the model that extracts the details takes minutes per document on a CPU, far too slow to run on every email.
Transit Ledger runs entirely locally on a CPU server. A fine-tuned 270M-parameter model screens each document first and discards everything that isn’t an invoice or receipt. Extracted totals are checked in code, and anything that doesn’t add up is flagged for review.
Screening takes about 0.4 seconds per document, so the slow extraction model only runs where it adds value. Real-world data also exposed a shortcut the first classifier had learned, which is why we keep testing beyond the first dataset.
Any document-heavy process, such as invoices, delivery notes, claims or applications, can use a small model as a gatekeeper: private data stays on your own infrastructure, and the expensive model only sees the documents that matter.
SPAM Escudo: reduce AI costs without sacrificing accuracy
Sending every contact-form submission to an LLM adds cost and latency, and most submissions are spam.
SPAM Escudo filters messages in five stages: schema validation, a honeypot, rules, embeddings and finally an LLM. The cheap stages only ever say “spam”; anything that might be genuine goes to the LLM. Every spam verdict from the LLM feeds new rules, and a weekly AI-assisted review finds ways to decide more messages earlier.
On the same 956 real messages, pre-LLM decisions rose from 54.9% to 78.1%, and past 90% with the full pre-LLM stack. That is 78% fewer LLM calls, with no genuine message wrongly flagged in that test.
Any high-volume AI workflow, such as support tickets, lead qualification or content moderation, can work the same way: resolve routine cases cheaply, reserve expensive AI for the exceptions, and watch costs fall over time.
Cerberus ID: automated decisions you can trust
Aletheia Systems’ Cerberus ID, for which IT-ISS is the technology architecture partner, scores college registrations to detect “ghost student” financial-aid fraud. Approving a fraudulent registration is expensive; sending every registration to manual review doesn’t scale.
A calibrated scoring engine, not an LLM. The business set the requirement first, at least 95% precision for automatic approval, and the thresholds were calibrated to meet it. The uncertain middle goes to review or to an automated ID and liveness check.
Calibrated against 10,806 real registrations, automatic approval reached 99.5% precision, with 78% fewer false approvals.
Wherever software approves, rejects or escalates, such as onboarding, claims, credit or compliance checks, thresholds can be set from what an error costs your business and proven against your historical data before go-live.
AI workflow design
In each of these projects, most of the value came from the design of the workflow, not the choice of model. We work with you to identify the processes worth automating, decide which steps need rules, a small model, an LLM or a person, and design an architecture that fits around your existing systems.
- A map of your process showing where AI adds value, and where it doesn’t
- A stage-by-stage design with the cost, speed and risk of each step
- Confidence thresholds and review routes based on what an error would cost
- A plan to benchmark the design on your real data before it goes live
Where AI agents fit: give them validated tools
Agents become useful when they can act through tools, and the Model Context Protocol (MCP) is now a common way to provide them. The business case is reliability: in a proof of concept at Runner Technologies, an agent validated addresses through a dedicated validation service instead of guessing. Don’t ask the model to judge whether an address is real; give it a tool that knows.
Five principles for putting AI into production
- Start with deterministic checks. Schema validation, lookups, arithmetic and business rules are fast, free and explainable. Use them before any model, and to check what the models produce.
- Gate expensive models with cheap ones. A small model answering one narrow question can decide what the large model ever needs to see.
- Set thresholds from the cost of an error before you tune anything, then calibrate against real data to meet them.
- Route uncertainty somewhere useful: a review status, a secondary check or a person. Never let it pass silently.
- Test on real-world data, and keep improving. The first dataset will not contain the cases that break your model, and every expensive decision should make the cheap stages better.
Have an AI project in mind?
Whether you want to automate an existing workflow, reduce what you spend on AI or bring AI into your business for the first time, we can help you work out what is practical. Tell us about your challenge, and we’ll discuss your requirements, explore possible solutions and agree the next steps.
Discuss your AI project →Or read about our approach to enterprise AI in production.