AI Agents in Production: Multi-Enterprise Implementation Strategies

canva_/kpmg/autodesk/_lightspeed · Tech · 2026 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise go-to-market automation

Multiple enterprises needed to move AI agents from pilots into production workflows across marketing, sales, data enrichment, and customer-facing operations.

Constraints:

governance
adoption
KPIs
data security
change management

Architecture

The panel’s shared production architecture is an enterprise agent platform layer, primarily Relevance AI, connected into go-to-market systems such as CRM, Slack, internal data, marketing/content workflows, and customer outreach. Companies use multi-agent workflows where traceability, explainability, human approval, RBAC, audit logs, integration scoping, sandbox rollout, and monitoring are part of the operating model. Autodesk, Canva, Lightspeed, and KPMG emphasize that adoption/governance is as important as model capability.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  business-workflow-selection["business-workflow-selection"]
  agent-platform["agent-platform"]
  multi-agent-workflows["multi-agent-workflows"]
  enterprise-integrations["enterprise-integrations"]
  human-approval-layer["human-approval-layer"]
  governance-controls["governance-controls"]
  adoption-operating-model["adoption-operating-model"]
  output-monitoring["output-monitoring"]
  business-workflow-selection --> agent-platform
  agent-platform --> multi-agent-workflows
  multi-agent-workflows --> enterprise-integrations
  multi-agent-workflows --> human-approval-layer
  agent-platform --> governance-controls
  enterprise-integrations --> governance-controls
  adoption-operating-model --> business-workflow-selection
  agent-platform --> output-monitoring
  output-monitoring --> governance-controls

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`business-workflow-selection`	Chooses bounded, revenue-adjacent, repetitive, or high-friction workflows for agentization.	`agent-platform`
`agent-platform`	Relevance AI-style platform for rapidly prototyping, running, observing, and governing multi-agent workflows.	`business-workflow-selection`, `multi-agent-workflows`, `human-approval-layer`, `enterprise-integrations`, `governance-controls`
`multi-agent-workflows`	Collaborating agents for research, outreach, enrichment, duplication, marketing, or customer-facing work.	`agent-platform`, `enterprise-integrations`, `human-approval-layer`
`enterprise-integrations`	Connections to CRM/Salesforce, Slack, marketing automation, internal data, and customer-facing systems.	`agent-platform`, `multi-agent-workflows`, `governance-controls`
`human-approval-layer`	Human-in-the-loop approvals for reports, emails, outputs, and staged rollout before autopilot.	`multi-agent-workflows`, `agent-platform`
`governance-controls`	RBAC, integration scoping, sensitive key protection, monitoring, observability, and audit logs.	`agent-platform`, `enterprise-integrations`
`adoption-operating-model`	Change-management, shared analytics/KPIs, AI leads/pods, and product-like ownership needed for rollout.	`business-workflow-selection`, `agent-platform`
`output-monitoring`	Monitors whether agent output remains aligned with intended behavior and quality standards.	`agent-platform`, `governance-controls`

Data flow

Teams identify a high-value but bounded workflow, prototype it in a sandbox on the agent platform, wire it to existing data/systems, and run outputs through human approval or review. Once outputs meet quality standards, they scale into production surfaces such as Slack, Salesforce/CRM, outreach, or internal workflows. Governance monitors users, integrations, audit logs, and output intent; adoption loops feed human feedback back into prompts, processes, and operating models.

4 supporting quotes

“we want to make sure that we start in a sandbox and we're able to really test and iterate before we roll something out”transcript, stated

“use some of relevance as functionality around human loop or us being able to really validate some of the outputs for different use cases”transcript, stated

“we can think about how to bring AI automation into existing workflows into existing surface area, whether that's surfacing them in Slack, surfacing them in our CRM in Salesforce as well.”transcript, stated

“And we have monitoring and observability audit logs. We need to be able to see what our users are doing within the system.”transcript, stated

Why this architecture

The companies chose platforms and staged operating models because enterprise value depends on fast commercial validation, traceable multi-agent collaboration, secure integrations, and adoption. Open-source frameworks offered capability but took too long for clients deciding commercial viability; UI point solutions added more surfaces for reps; and unguided pilots failed when adoption/change management was ignored.

4 supporting quotes

“Crew AI um and then moved on to things like Autogen um when that was emerging. Um and really what we found was that whilst those had great capabilities, great depth of capability, um they took too long to actually service the client need”transcript, stated

“clients don't want to wait multiple weeks or or months to work out whether something's going to have a commercial return.”transcript, stated

“being traceable and being explainable were the two main kind of assessment factors for us when we were going through this process.”transcript, stated

“90% of what we want to do with agents the technology is already there and so that the technical aspect is probably there. It's how do we actually get adoption within the business”transcript, stated

Models used: LLM agents

Scale

Autodesk reports tasks that previously took hours could be reduced to minutes or seconds with agents.
“things that would take hours on me to create like content that is suited for onetoone presentation it that will take hours but with AI agents it will take minutes if not seconds”transcript, stated
Canva needed controls that let the platform move from one controlled use case to many teams.
“foundational security controls that enable us to move from one use case that it is controlling everything to being able to onboard lots and lots of teams onto a platform”transcript, stated
Lightspeed’s closed-lost research/outreach agent operates in a context with hundreds of global sales reps and a diverse tech stack.
“Lightseed has hundreds of sales reps globally and also an amazing and diverse tech stack. With this I think comes a lot of complexity.”transcript, stated
Canva highlights small quick wins such as saving reps 15 minutes per day to build momentum.
“help help our reps, you know, save 15 minutes every day prepping for a call with a new prospect or um write that follow-up email.”transcript, stated

Alternatives considered (roads not taken)

Open-source agent frameworks as the main client-delivery path

Rejected because: They had depth but took too long to prove commercial viability for clients.

“whilst those had great capabilities, great depth of capability, um they took too long to actually service the client need that we had in front of us”transcript, stated

Add another standalone UI for customer-facing teams

Rejected because: Teams already use many platforms, so agents should surface inside existing Slack/CRM workflows.

“as a customerf facing person like seven to eight different platforms. And so we're really excited about platforms like relevance because we can think about how to bring AI automation into existing workflows”transcript, stated

Immediate autopilot for outbound/customer-impacting agents

Rejected because: Lightspeed used crawl-walk-run approval gates and reviewed reports/emails/replies until quality was trusted.

“initially we approved all the research reports then we approved all the emails and then we approved all the email replies until we were fully comfortable that the agent was performing exactly how we wanted to.”transcript, stated

Patterns

`multi-agent-orchestration` (gtm)

The case describes enterprise multi-agent systems across several revenue and knowledge-work workflows.

“They work with enterprise clients to pilot and prototype AI agents quickly, particularly for complex multi-agent systems that clients struggle to implement themselves.”summary, stated

`eval-driven-development` (operations)

Teams emphasize KPIs, feedback, iteration, and production adoption rather than one-off demos.

“h marketing operations sales operations BD and sales to get their feedback iterate but also uh making sure that uh we have clear KPIs that cross”transcript, stated

Tradeoffs

How to roll out agents for customer-impacting work.

Chose Crawl-walk-run with human approvals before autopilot. over Immediate full automation of outreach and customer-facing outputs..

Rationale: The transcript describes approving research reports, emails, and replies until the agent matched expectations.

“we followed a crawl walk run approach. We relied really heavily on relevance as human in the loop um approvals functionality both through the platform and through Slack.”transcript, stated

“We didn't put it on autopilot. So that process of continuous iteration and refinement was really important.”transcript, stated

How to integrate agent outputs into go-to-market work.

Chose Surface automation inside existing systems like Slack and Salesforce. over Adding yet another point-solution interface to reps’ daily workflow..

Rationale: Canva noted customer-facing teams already use many platforms, so hidden automation in existing surfaces reduces adoption friction.

Gotchas (what bit them)

Scaling agents requires adoption and governance work, not only model capability.

Root cause: Different teams need feedback loops, KPIs, and safe integration with enterprise systems.

Resolution: Treat agents as change-management programs with metrics and stakeholder iteration.

“The companies faced challenges around scaling AI from proof-of-concept to production, managing agent quality and accuracy, and driving adoption across diverse teams.”summary, stated

Evaluation

Organizations report time savings, pipeline, engagement, onboarding, and departmental scale-out metrics.

Metrics: hours to minutes · pipeline generation · engagement rates · onboarding speed

“Results include significant time savings (tasks taking hours reduced to minutes), improved pipeline generation, increased engagement rates, faster customer onboarding, and the successful scaling of AI agents across multiple department”summary, stated

Outcomes

The panel reports production adoption across multiple departments and companies.

Multi-enterprise platform series, likely promotional toward Relevance AI.

“And with this series, we had one goal, bring you content from the world's top companies on how they're winning with AI agents.”transcript, implied