New Odyssey
Stuck as human middleware? Autonomous operators and New Odyssey
Autonomous operators are in the funding news. How ops teams can judge any AI agent approach on trust, reporting and reach across existing tech stacks.
Ademola Afolabi
Founder
Operations teams stuck as human middleware between their systems are hearing a lot of funding news about autonomous operators, and that news shows investors believe AI agents can take on back-office work. It doesn't tell a COO whether an agent will stop staff re-keying data between ERP, HRIS, CRM, ticketing and document systems, or whether security and audit will trust it. Judge any agent approach, funded or not, on three questions. Does it work across your existing tech stack? Is every action audited and explained, with workflow approvals? Will you get a working version on real systems, fixed in scope, with an honest go or no-go?
Why your team is still re-keying data, whatever the funding headlines say
Your team is still re-keying data because the work sits in the gaps between your systems, and no single tool owns those gaps. Operations staff spend hours every week copying the same records from the ERP into the HRIS, from the CRM into ticketing, and from email attachments into the document system. Every hand-off adds delay and a chance of error.
The people doing that copying have become human middleware. Each platform in the stack was bought to do one job well. The re-keying, document handling, reconciliation and month-end close tasks between them fall to whoever is nearest.
Human middleware doesn't stay the same size. Claims, onboarding and invoicing volumes keep climbing while headcount stays flat, so the work between systems grows with them. Backlogs lengthen and customers wait longer for decisions.
A funding round for an autonomous operators company is a signal that the market sees this problem. It is not proof that any one product solves it for your existing tech stack, your approval rules or your audit requirements.
What are 'autonomous operators' and AI agents for operations, in plain terms?
Autonomous operators, or AI agents for operations, are software agents that carry out back-office tasks rather than only answering questions. Examples include reading an invoice and matching it to a purchase order, processing a new-hire packet, or routing a claim.
The label matters less than three practical details:
- Where the agent can read and write. Can it pull from your ERP and write the result back into your HRIS or CRM, or does it only work inside one vendor's product?
- Who approves its actions. Are there workflow approvals and segregation of duties, or does the agent act unchecked?
- What reporting shows afterward. Can someone see what the agent did, when, and why?
Before you rely on any headline, Evos included, read the company's own announcement and product documentation. Funding coverage often describes ambitions. Your evaluation needs to rest on what the product does today with systems like yours.
Why do AI pilots stall before they reach production?
AI pilots stall because security, compliance and audit won't sign off on agents whose actions and decisions they can't trace or explain. A pilot that impresses in a demo can still fail review if no one can answer basic questions. What did the agent change? Who approved it? Which version was running? Why did it make that decision?
Pilots built without audit trails, workflow approvals and versioned deployments can't clear that review, so they never scale past the test group. Adding the controls afterward often means rebuilding the pilot.
The second reason is ROI. Leadership asks for proof that AI spend is paying off, but most experiments never produce a measured before-and-after on a real back-office process. Without a baseline (volume, handling time, error rate), there is nothing to compare against, and the business case stays a guess.
Comparing the approaches: vendor-native agents, integration platforms, consultancies and agent providers
Each common approach solves part of the problem. The question is which part it leaves with your team.
Vendor-native agents are built around their own platform's data and workflows. A CRM vendor's agent is useful inside the CRM, but it stops at the edge of that one system, which is exactly where the work between systems begins.
Integration platforms sell the tooling to connect systems. Design, build and maintenance stay with your team or an implementation partner, along with any governance layer for AI agents.
Traditional consulting sells time. Months of discovery billed at day rates often come before anything runs, and the output is frequently a slide deck rather than a working workflow with a committed outcome.
Newly funded agent startups should be judged by the same criteria as everyone else. The size of a round says something about investor confidence. It says nothing about whether the product reaches your existing tech stack or will pass your audit review.
| Approach | Reach across your existing tech stack | Governance and trust | Time to a working version | Who maintains it |
|---|---|---|---|---|
| Vendor-native agents | Mostly within one platform | Depends on the vendor's controls | Fast inside that platform | The vendor, within its platform |
| Integration platforms | Broad, if your team builds it | Your team builds it | Depends on internal capacity | Your team or a partner |
| Traditional consulting | Varies by engagement | Varies; often a recommendation, not a running control | Often months of discovery first | Often handed back to you |
| Newly funded agent startups | Check against your stack | Check for audit, versioning and approvals | Check for a pilot on real systems | Check the support model |
| Governed cross-platform agent providers | Built to span many platforms | Controls in place from the start | Days to weeks | Provider, with managed options |
What does 'good' look like? A self-assessment for any AI agent approach
A good AI agent approach works across your systems, earns the trust of your risk teams, and proves its economics on one process before you scale it. Score any vendor, internal project or proposal against these six questions:
- Cross-system reach. Does the agent work the same whichever HRIS, ATS, accounting and CRM vendors you run, and does it write output back where it belongs so nobody re-keys it?
- Trust and governance. Is every agent action audited, every deployment versioned and every decision explained in terms a non-engineer can read?
- Workflow approvals and segregation of duties. Can your risk team see and enforce who approves what, and can the same person both trigger and approve an action?
- Exceptions. When the agent can't handle a case, does it go to a named owner, or get skipped without anyone knowing?
- Measurement. Will you get a baseline cost for one process, a working version on real systems, and a clear go or no-go decision?
- Scope. Is each stage fixed in scope before it starts, so you know what you are buying and what "done" means?
If an approach answers "no" or "it depends" to questions 2 through 4, expect it to stall at security review. If it answers "no" to question 5, expect to have a hard time defending the spend to leadership.
Could we just build it internally?
Building AI agents internally is a valid option, especially for teams with strong engineering capacity and a stable stack. The trade-off is that your team then owns connector design, build and ongoing maintenance as each vendor changes its APIs, plus the governance layer that lets the agents into production.
To clear security and compliance review, an internal build needs to cover at least:
- An audit log of every agent action and decision
- Versioned deployments, so you know which logic was running at any point
- Explainable decisions, recorded in human-readable terms
- Workflow approvals and segregation of duties
- Exception routing to a named owner
One useful reference point: New Odyssey's governance control plane, CIP, is open source under the MIT license. Whether you build internally or buy, your team can inspect how one governance approach handles auditing, versioning and decision records.
How do you get security and compliance to trust AI agents?
You get security and compliance to trust AI agents by bringing them in at the start and showing them controls they can inspect, not promises. Risk teams that first see an agent after the pilot tend to find gaps that send the project back to the beginning.
The controls risk teams typically need to see are:
- An audit log of every agent decision
- Workflow approvals at the points where an agent's output changes a record, a payment or a decision
- Segregation of duties, so no single person or agent both initiates and approves
- Versioned deployments, so any past action can be traced to the exact logic that produced it
As one example, New Odyssey runs its agents on CIP, which audits each action, versions each deployment, records each decision with a human-readable explanation, and routes exceptions to a named owner rather than skipping them. Those records give risk teams the reporting they need to approve a rollout instead of blocking it.
Is a smaller or less-known provider a risk?
A smaller provider is a risk only if its delivery model and evidence don't hold up, and the same is true of a large or heavily funded one. Judge any provider on whether it can show a working version on your real systems, whether its controls pass your review, and whether each stage is fixed in scope.
For New Odyssey, the basic facts are these. The company was founded in London in 2016, is registered in England and Wales (company 10193156), and has US operations in Dallas. It works with mid-market and enterprise operations, finance, HR, legal and compliance teams. You can check the company background on the About Us page at new-odyssey.com.
A realistic path: from one workflow to governed rollout
The realistic path starts with one high-volume process, measured before anything is built, and scales only when the numbers justify it.
- Pick one process. Choose something with real volume, such as invoice reconciliation, employee onboarding or claims intake.
- Measure the baseline. Record volume, handling time and error rate, so there is a before to compare against.
- Build a working version on real systems. Test it against live data and your actual approval rules, not a sandbox.
- Decide go or no-go. Compare the results with the baseline. A recommendation not to proceed is a legitimate outcome when the numbers don't hold up.
- Roll out with governance and support. Harden for production, then keep the workflow working as systems, rules and volumes change.
New Odyssey's version of this path runs in fixed stages: a fixed-fee Process Assessment, a 10-business-day Automation Sprint on real systems, then Production Implementation and Managed Automation. Agents connect to 200+ HRIS, ATS, accounting, CRM, ticketing and file-storage platforms through Merge's unified APIs.
Timelines depend on the route:
- Packaged Pegasus agents (Compliance, Finance, Contract and People) start with a pilot on one real workflow, and most teams are live within 3–5 business days.
- Pegasus suites across a whole HRIS, ATS, accounting or CRM category take four weeks from connecting systems to go-live, followed by 30 days of hypercare.
- Enterprise programs on Phoenix, for workflows no packaged agent covers, run from a 10-Day Agent Sprint through production hardening to managed operations and typically reach production in about six weeks.
Next step: check your own workflow against the checklist
Pick one process your team owns, the one with the most re-keying or the longest backlog, and score it against the six questions in the self-assessment above. Note where your current tools or proposals answer "no" or "it depends."
If you'd like to talk that workflow through, use Assess my workflow on the New Odyssey homepage to discuss your situation, including whether automation is the right answer at all.
Frequently Asked Questions
- Does a large funding round mean an AI agent vendor will work for our operations team?
- No. A funding round signals investor confidence, but it doesn't show whether the product works across your existing tech stack or will pass your security and audit review, so judge every vendor on reach, governance and a working version on real systems.
- Why do AI agent pilots fail to reach production?
- Most AI agent pilots stall because security, compliance and audit can't trace or explain what the agent did, and because the pilot never produced a measured before-and-after on a real process. Audit trails, workflow approvals and versioned deployments need to be in place from the start.
- What controls do risk teams need to see before approving AI agents?
- Risk teams typically need an audit log of every agent decision, workflow approvals, segregation of duties, versioned deployments and exception routing to a named owner. Those controls give them the reporting to approve a rollout.
- How long does it take New Odyssey to get an AI agent live?
- New Odyssey's packaged Pegasus agents start with a pilot on one real workflow, and most teams are live within 3–5 business days. Pegasus suites take four weeks plus 30 days of hypercare, and enterprise programs on Phoenix typically reach production in about six weeks.
- Can we inspect New Odyssey's AI governance approach before buying?
- Yes. New Odyssey's governance control plane, CIP, is open source under the MIT license, so your team can inspect how it handles auditing, versioning and decision records.
Published
November 18, 2026
Last updated
November 18, 2026
Author
Ademola Afolabi · Founder
Reviewer
New Odyssey Editorial Team · Editorial Review
Methodology
This new odyssey piece is maintained by the New Odyssey editorial and delivery teams using current implementation patterns and operating assumptions.
Which of your workflows is worth automating?
Answer three questions about the process you want to automate and see what it costs you today.
Assess my workflowStay updated
Get weekly integration insights delivered to your inbox.