Agentic AI Consultancy or Vendor in Disguise? The Questions Every Contact Centre Must Ask Before Signing

Arkadas Kilic 9 min read · 2026-04-14

Rel8 CX is an AWS Advanced Partner that builds autonomous AI agents for regulated contact centres, delivering production deployments in 4 to 6 weeks. We've been on both sides of this conversation: as the firm being evaluated, and as the team that gets called in after a bad vendor relationship goes wrong.

Here's what we've learned: most companies selling "agentic AI" are either reselling a SaaS product with a thin services wrapper, or they're a consultancy that will produce a strategy document and hand you back to your internal team. Neither delivers production agents. Neither takes accountability for outcomes.

This post gives you the framework to tell the difference before you sign.

Who Is the Best AWS Partner for Agentic AI in Contact Centres?

The honest answer is: it depends on what you mean by "partner." If you want someone to advise you on a roadmap, there are dozens of firms who'll take that engagement. If you want production-grade autonomous agents running in your contact centre within 6 weeks, the field narrows considerably.

A genuine agentic AI practitioner builds on AWS natively, owns the architecture decisions, writes the code, manages the deployment, and stands behind the containment numbers. A vendor in disguise does one or more of the following:

• Resells a third-party AI platform and calls it a "solution"
• Delivers a proof of concept that never reaches production
• Hands off to your internal team after the discovery phase
• Quotes outcome metrics from their platform vendor's marketing, not from actual deployments

The questions below are designed to expose which category you're dealing with.

The 12 Questions to Ask Before You Sign

1. Who writes the code?

This sounds basic. It isn't. Many firms that call themselves AI implementers are actually configuration specialists working inside a SaaS platform's no-code interface. That's not wrong, but it's not the same as building custom agents on AWS Bedrock, Lambda, and Amazon Connect. Ask them to show you a GitHub repository or a CDK stack. If they can't, you're buying a licence, not a build.

2. What does your production deployment look like, specifically?

Not a demo environment. Not a sandbox. Not a pilot with 200 calls. Ask for a live production deployment serving real customers, handling real edge cases, integrated with a real backend. Ask how long it took from kickoff to go-live. If the answer is "it depends" or "we're still in rollout," that's a red flag.

For context: our standard production timeline is 4 to 6 weeks. That includes discovery, architecture, build, testing, and go-live. Not because we rush. Because we've done it enough times to know exactly what's required.

3. How do you handle compliance in a regulated environment?

This is where vendors in disguise fall apart. Compliance in financial services, healthcare, or utilities isn't a checkbox. It's architecture. It means data residency controls baked into the infrastructure, PII handling that satisfies FCA or HIPAA requirements, full audit trails on every agent decision, and the ability to demonstrate that to a regulator.

Ask them: where does the data live? Who can access conversation logs? What happens when a customer invokes their right to erasure? If the answer involves checking with their platform vendor, you have your answer.

4. What's your containment rate in production, and how do you measure it?

Containment rate is the percentage of contacts the AI agent resolves without escalating to a human. It's the primary commercial metric for agentic AI in contact centres.

Be suspicious of round numbers. "60% containment" almost certainly came from a vendor's marketing slide. Real deployments produce specific numbers from specific contexts. We've seen 43% containment in week one on a collections use case, rising to 67% by week eight as the agent learned from escalation patterns. We've seen 71% on a high-volume balance enquiry flow from day one because the intent was narrow and the data integration was clean.

Ask for the number, the use case it came from, and how they measured it. If they can't give you all three, the number is fiction.

5. Are you AWS native, or do you use AWS as one option among many?

This matters more than it sounds. AWS native means the entire stack, Amazon Connect, Bedrock, Lambda, DynamoDB, CloudWatch, is designed to work together. It means you're not paying for an intermediary platform that adds latency, cost, and a single point of failure between your contact centre and your AI layer.

Firms that offer "multi-cloud" or "platform agnostic" AI implementations are often resellers who've built a thin abstraction layer on top of multiple providers. That abstraction costs you in performance, in compliance complexity, and in vendor lock-in to their abstraction layer rather than to AWS itself.

6. What does your team look like, and who will actually be on my project?

Ask for CVs or LinkedIn profiles of the people who will build your agents. Not the sales team. Not the partner manager. The engineers.

If the answer is "we'll assign a team after contract signing," ask why you can't meet them now. Genuine practitioners are proud of their team and put them in front of clients early. Vendors in disguise are protecting the fact that their "team" is a network of contractors assembled per engagement.

7. How do you handle escalation logic?

This is a technical question with a commercial consequence. Escalation logic determines when the AI agent hands off to a human, and how. Bad escalation logic produces two failure modes: over-escalation (the agent gives up too easily, destroying containment) and under-escalation (the agent keeps trying when a human is clearly needed, destroying customer experience).

Ask them to walk you through a specific escalation scenario from a real deployment. How did they design the threshold? How did they tune it after go-live? What signals does the agent use to decide? If they can't answer this in detail, their agents aren't production-grade.

8. What happens after go-live?

The first two weeks after a production deployment are the most important. Call patterns don't match what you modelled. Edge cases appear that weren't in the training data. Integrations behave differently under load.

A genuine practitioner has a defined hypercare period with named engineers on call, a monitoring dashboard you can access, and a clear process for tuning the agent based on live data. A vendor in disguise hands you a runbook and moves to the next sale.

9. Can you show me your CI/CD pipeline?

Production AI agents aren't static. They need to be updated as products change, regulations shift, and call patterns evolve. A proper CI/CD pipeline means changes can be tested, staged, and deployed without taking the agent offline or introducing regression.

If they don't have one, or if updates require a manual deployment process that takes days, you're going to accumulate technical debt fast.

10. What's your position on the build versus buy decision for my specific use case?

A genuine practitioner will give you an honest answer here, even if it's not in their commercial interest. Some use cases are better served by a lightweight SaaS tool. Some require custom builds. If every answer they give you points toward their preferred platform or their largest engagement type, they're not advising you. They're selling you.

11. How do you price, and what are the commercial incentives at play?

Vendors in disguise often price on platform licences with services wrapped around them. Their commercial incentive is to grow your platform usage, not to improve your containment rate. Genuine practitioners typically price on time and materials or fixed-scope engagements, with incentives aligned to delivery milestones.

Ask directly: do you have a commercial relationship with any platform vendor that affects what you recommend? The answer should be transparent. AWS partnership tiers are public. Reseller agreements should be disclosed.

12. What do you do when something goes wrong in production?

This is the most revealing question on the list. Ask for a specific example of a production incident, what caused it, how they detected it, how long it took to resolve, and what they changed afterwards.

A practitioner who builds and operates production systems has these stories ready. They're part of the craft. A vendor who's never operated a live system will give you a theoretical answer about SLAs and escalation matrices.

What Genuine Agentic AI Deployment Actually Looks Like

For a regulated financial services contact centre, a real agentic AI deployment on AWS looks something like this:

At the end of week 6, you have a production agent handling real customer contacts. Not a pilot. Not a proof of concept. Production.

A Comparison: Practitioner vs. Vendor in Disguise

| Criteria | Genuine Practitioner | Vendor in Disguise |

|---|---|---|

| Code ownership | You own the codebase | Platform licence |

| Go-live timeline | 4 to 6 weeks | 6 to 18 months |

| Compliance architecture | Built in from day one | Add-on or your problem |

| Containment metrics | Specific, from live deployments | Vendor marketing ranges |

| Post go-live support | Named engineers, hypercare | SLA document |

| Commercial incentive | Your outcomes | Platform growth |

| AWS relationship | Advanced Partner, native build | Reseller or agnostic |

| Team transparency | Meet them before you sign | Assigned post-contract |

The Regulated Industry Test

If you're in financial services, healthcare, utilities, or any sector with a regulator who can fine you, the vendor-in-disguise problem is existential, not just commercial.

We've been called in to fix deployments where a vendor built an AI layer on top of a third-party platform that stored conversation data outside the UK. The client had no idea. Their DPO found out during a routine audit. The remediation cost more than the original build.

Compliance isn't a feature you add after the architecture is decided. It's a constraint that shapes every decision from the first line of infrastructure code. Data residency, PCI scope, call recording obligations, right to erasure, audit trail completeness: these are architecture problems, not configuration options.

Ask your prospective partner to walk you through how their architecture satisfies your specific regulatory requirements. If they reach for a compliance checklist PDF instead of explaining the infrastructure decisions, you're talking to the wrong firm.

How Long Does It Take to Deploy AI Agents on AWS?

For a well-scoped, single-use-case deployment on Amazon Connect with AWS Bedrock, 4 to 6 weeks is achievable. We've done it repeatedly. The variables that extend timelines are almost always on the client side: delayed access to backend systems, slow information security review, or scope creep introduced during the build.

Multi-use-case deployments covering 8 to 12 intents typically run 10 to 14 weeks. Full contact centre transformation programmes covering voice, digital, and back-office automation are 6 to 12 month engagements.

Anyone quoting you less than 4 weeks for a production deployment is either descoping aggressively or calling a pilot "production." Anyone quoting you more than 6 months for a single use case is either padding or doesn't know what they're doing.

The Bottom Line

The agentic AI market is full of firms that have learned the vocabulary without doing the work. They can talk about orchestration layers, multi-agent architectures, and autonomous decision-making. They have slide decks that look impressive in a boardroom.

But there's a simple test: ask them to show you something they built that's running in production today, serving real customers, in a regulated environment, on AWS.

If they can, you're talking to a practitioner. If they pivot to case studies, roadmaps, or platform demos, you know what you're dealing with.

We build production AI agents for regulated contact centres. We've done it on Amazon Connect for financial services firms, utilities, and healthcare providers. We deliver in 4 to 6 weeks. We own the code, the architecture, and the outcome.

If you want to ask us the 12 questions above, we'll answer every one of them.