What a company-wide AI agent actually does
Plans and executes multi-step work across an enterprise's CRM, ERP, ITSM and data platforms, routing anything consequential to a human and logging every action for audit.
An enterprise AI agent is less a single chatbot and more an orchestration layer: a planning model breaks a request into steps, calls tools across whatever systems of record the company runs (Salesforce or SAP, ServiceNow, a data warehouse), and either completes the task, drafts it for a human to approve, or hands it off entirely. At this scale the agent itself is rarely the hard part; the hard part is governance: who can grant an agent write access to a production system, how actions get logged for audit, and how the company proves which agent did what when something goes wrong.
Gartner expects 40% of enterprise applications to carry a task-specific AI agent by the end of 2026, up from under 5% in 2025, and agentic AI spending is projected to hit roughly $202 billion in 2026. But adoption and deployment aren't the same thing: the same research found only 17% of organizations have actually deployed an agent, and McKinsey's 2025 State of AI survey put the share scaling an agentic system in any single function at 10% or less.
Where enterprise agents still fail: Gartner estimates that of the thousands of vendors now claiming 'agentic AI,' only about 130 are offering something real, and it separately predicts over 40% of agentic AI projects will be canceled by the end of 2027 over cost, unclear value and weak risk controls. The agents most likely to survive that cull are narrow, measurable and reversible: a single process with a human still able to override every consequential decision, not a company-wide rollout on day one.
AI capabilities involved
Planning and decomposing a multi-step request into tool calls
Routing and classification at high volume across departments
Acting across CRM, ERP and ITSM systems through tool calls
Summarizing and logging every action for audit and compliance
Who uses this
- IT and platform teams evaluating Microsoft, Google or AWS agent infrastructure before a company-wide rollout
- Enterprises already running Salesforce, ServiceNow or SAP deciding whether to turn on the native agent or build custom
- Security and compliance teams scoping governance for agents with write access to production systems
- Operations leaders piloting one high-value process (support, procurement, IT helpdesk) before a broader commitment
What the enterprise agent handles (and what it shouldn't)
| Task | Agent handles it? | Systems it touches | Human role |
|---|---|---|---|
| Answering employee IT and HR helpdesk questions | Fully | ITSM (ServiceNow, Atomicwork), HR system | None for documented policy |
| Resetting access or provisioning standard software | With approval | Identity provider, ITSM | Approves anything touching production access |
| Drafting customer support replies at high ticket volume | With approval | CRM/helpdesk, knowledge base | Reviews anything escalated |
| Writing to a production CRM or ERP record (price, contract term) | No | CRM/ERP | Makes the decision and the write |
| Summarizing meetings, tickets or documents across departments | Fully | Document stores, meeting tools | None |
| Procurement intake and supplier follow-up | With approval | Procurement platform, ERP | Approves anything above a spend threshold |
| Security alert triage and first-pass investigation | With approval | SIEM/EDR | Confirms and responds to real incidents |
Off-the-shelf AI agents on the market
Real products you can sign up for today — with current 2026 pricing, honest pros and cons.
Salesforce Agentforce
Enterprises already running Salesforce Service or Sales Cloud
Salesforce Foundations: a free starting point with Agent Builder included
$2 per customer-facing conversation, or Flex Credits at $500 per 100,000 credits (a standard action costs 20 credits, a voice action 30)
Agentforce 1 / Max edition from $550/user/mo, including 2.75M Flex Credits per org per year
Pros
- +Works directly on your existing CRM and Data Cloud data without a separate integration layer
- +Add-on user licenses from $5/mo make it cheap to pilot on one team
- +Folding in Fin's customer-service agents after the 2026 Salesforce acquisition
Cons
- −Requires the Salesforce ecosystem, so it's not a fit if you're not already there
- −Setup is estimated at $2K-$6K per agent by third-party implementers, on top of licensing
Sierra
Large enterprises willing to run a full outcome-based vendor contract
None public; enterprise sales only
Contact sales; reported starting contracts around $150,000/yr with outcome-based pricing
Negotiated per-resolution rate at scale; escalations to a human are free
Pros
- +Purpose-built 'Agent OS' for enterprise voice and chat, not a bolt-on to an existing suite
- +Can take actions in back-office systems, not just answer questions
- +Backed by a $950M raise at a $15.8B valuation in May 2026, with roughly $200M in reported ARR
Cons
- −No public pricing; procurement alone can take months
- −Enterprise-only, with long deployment timelines
Decagon
Enterprises standardizing customer-facing agents across multiple channels at once
None; demo only
Contact sales; third-party estimates report roughly $50,000/yr platform fee plus about $0.99 per conversation
Custom; reached a $4.5B valuation on a $250M Series D in January 2026
Pros
- +'Agent Operating Procedures' let you write natural-language SOPs instead of flowcharts
- +Covers chat, email and voice from one platform
- +Enterprise-grade investment and scale behind it
Cons
- −No public pricing page at all
- −Enterprise-only; not a fit for piloting a single small process cheaply
ServiceNow AI Agents (Now Assist)
Enterprises already running ServiceNow for IT, HR or customer service workflows
None public
Contact sales; third-party estimates put it around $90-$200/user/mo plus assist usage, restructured into Foundation/Advanced/Prime tiers in April 2026
Only the Prime tier can build net-new custom agents
Pros
- +Deep native fit if you already run ServiceNow for ITSM
- +Includes the former Moveworks employee-support agent after ServiceNow's roughly $2.4B acquisition closed in December 2025
- +AI Control Tower gives IT a single place to govern every agent on the platform
Cons
- −Very opaque pricing with heavy implementation requirements
- −Only the top Prime tier unlocks building your own agents, not just using prebuilt ones
The AI stack
At enterprise scale the stack decision is really a platform decision: which cloud already hosts your systems of record, and how you govern what the agent is allowed to do there.
Reasoning model
Plans multi-step work and decides what's safe to execute without approval
Claude Opus 5.5
$4/M input, $20/M outputThe planning/orchestration layer
GPT-6 Astra
$10/M input, $50/M outputTeams already standardized on OpenAI for the hardest tasks
Claude Haiku 5.5
$0.10-$0.50/M input tokensSub-agents handling routing and classification
Our pick: Use a strong planning model (Opus 5.5 or GPT-6 Astra) only at the orchestration layer, and route high-volume classification work to Haiku 5.5 or GPT-6 Luna to control cost.
Platform & hosting
Where the agent actually runs, how it authenticates to your systems, and who hosts the sandbox
Microsoft Foundry Agent Service
$0.0994/vCPU-hr + $0.0118/GiB-hr, active sessions onlyMicrosoft-standardized enterprises
AWS Bedrock AgentCore
$0.0895/vCPU-hr + $0.00945/GB-hr runtime, plus Gateway and Memory feesAWS-standardized enterprises with an internal platform team
Gemini Enterprise Agent Platform (Agent Runtime)
Enterprise plans, contact salesGoogle Cloud-standardized enterprises
Our pick: Pick the platform that matches where your systems of record already live; switching clouds to chase a marginally better agent runtime rarely pays off.
Security & governance
Controls what an agent is allowed to touch and proves what it did after the fact
OWASP Agentic AI Top 10 program (ASI01-ASI10)
Internal process, no license feeEvery enterprise deployment, regardless of platform
Vendor-native governance (ServiceNow AI Control Tower, Sana Agent System of Record)
Included with the platformEnterprises consolidating on one platform
Our pick: Apply least-privilege scoping and human approval on every consequential write before your first agent touches a production system, not after.
Observability & evals
Catches drift, hallucinated actions and runaway tool calls before they become an incident
Langfuse
Free up to 50K units/mo, then $29-$2,499/moCustom and hire-agency builds
LangSmith
Free (5K traces), Plus $39/seat, deployments metered in LSUs at $1 eachTeams standardized on LangChain
Our pick: Wire in tracing and weekly transcript review from day one; LangChain's 2026 State of Agent Engineering survey found only 52% of teams running agents in production actually run offline evals.
Reference architecture
The hardest engineering problem at enterprise scale is excessive agency: deciding exactly which actions an agent can take without a human in the loop, because a single overreaching write to a production CRM or ERP record is the kind of incident Forrester predicts will cause a publicly disclosed, agent-driven breach in 2026.
Request or trigger arrives
Chat, ticket, or scheduled jobAn employee or customer request, or a scheduled process, kicks off the agent
Plan the steps
Planning/orchestration modelA strong reasoning model (Opus 5.5 or GPT-6 Astra) breaks the request into a sequence of tool calls
Check permissions
Identity and access layerEach planned action is checked against what this agent is actually allowed to touch
Call tools across systems
CRM/ERP/ITSM connectorsThe agent reads and, where approved, writes to the relevant systems of record
Human approval gate
Approval workflowConsequential or irreversible actions pause for a named human to approve before executing
Execute and confirm
Target systemApproved actions execute, and the result is confirmed back to the requester
Log every step
Observability platformEvery plan, tool call and approval is logged for audit, with weekly sampling for drift
Escalate failures
Human queueAnything that errors out or falls below a confidence threshold routes to a named owner
Estimated cost per request
~$0.50-$3.00 per completed multi-step task in model costs alone, before any platform per-conversation or Flex Credit fee
Cost calculator
Drag the sliders to model your actual usage. The numbers update in real time so you can stress-test economics before writing a single line of code.
Estimate your monthly model cost for one or more automated processes.
Estimated monthly cost
$6,429
≈ $77.1k per year
Calculator notes
- processes_automated is contextual: each additional process roughly multiplies the model-cost line, since governance and observability tooling doesn't need to be rebuilt per process.
- The $1.20/task model cost assumes a planning step on Opus 5.5 or GPT-6 Astra plus several cheaper tool-call steps on Haiku 5.5 or GPT-6 Luna; a simpler single-step task costs closer to $0.05-$0.10.
- The platform per-conversation line shows what you'd pay buying a SaaS agent platform instead; most enterprises pick one path, not both.
- Enterprise platform contracts (Sierra, Decagon, ServiceNow Prime) are typically negotiated annually and aren't captured well by a per-task calculator; see build_vs_buy.
Build it yourself with vibe-coding tools
Microsoft Copilot Studio lets an enterprise team with Microsoft 365 and Azure already in place stand up a governed custom agent without starting a framework from scratch.
Time to MVP
2-6 weeks for a first working agent inside an existing Microsoft tenant; months to reach company-wide governance maturity
Total cost to MVP
$200-$1,000/mo in Copilot Studio credits (capacity packs of 25,000 credits at $200/mo) plus engineering time
You'll need
Starter prompt
Create a new custom agent in Copilot Studio scoped to one process, such as internal IT helpdesk triage. Set its instructions to: 'You are an internal IT support agent for {Company}. Read incoming tickets from the connected ITSM system and classify them as password/access, hardware, software or other. For password and access issues, check the knowledge base for the documented self-service steps and reply with them directly. For anything involving granting new access, elevated permissions, or touching a production system, do not act: create a ticket note summarizing the request and assign it to the IT team queue for human approval. Log every classification and action you take, including tickets you only summarized without acting on, so the IT team can audit your decisions weekly.' Connect the agent to your ITSM system (ServiceNow, Jira Service Management) as a knowledge source and an action connector, and connect your identity provider as a read-only lookup so the agent can confirm who is asking without being able to grant access itself. Set an approval step in the agent flow for any action tagged as touching production access, routing it to a named IT lead in Teams. Test it against your last 20 real tickets before enabling it on live traffic.
Paste this into Microsoft Copilot Studio
Follow-up prompts (run in order)
- 1
Add a second connector so the agent can also check hardware inventory status before escalating a hardware request
- 2
Add a weekly summary flow that emails IT leadership every ticket the agent handled and every one it escalated
- 3
Add a confidence threshold so tickets the agent isn't sure about always escalate, even if they look like password resets
- 4
Extend the same pattern to a second process, like HR benefits questions, as a separate scoped agent rather than widening this one
Expected output
A single governed agent handling one internal process end to end, with every write-worthy action routed through a named human approver and a weekly audit log.
Known gotchas
- !Copilot Studio's credit-based billing is hard to forecast until you've run a real month of traffic; a capacity pack can run out mid-month during a usage spike
- !Scoping creep is the most common failure: teams start with one process and widen the agent's permissions before the governance work catches up
- !Connecting to legacy on-prem systems often needs a separate Azure gateway or data connector, which adds setup time the pricing page doesn't account for
- !Without an explicit instruction to log every decision, the agent's reasoning for escalations and approvals isn't visible after the fact
- !Enterprise IT security review of a new agent's permissions frequently takes longer than building the agent itself
- !A single company-wide rollout without a pilot is exactly the pattern Gartner's cancellation research flags as high-risk
Compliance & risk reality check
Enterprise agents carry heavier governance weight than a small-business or single-function agent, because the blast radius of a mistake is bigger and the regulatory surface is wider.
Consequential-decision rules
Colorado's SB 26-189 (effective Jan 1, 2027), California's CCPA ADMT duties (from Jan 1, 2027) and the EU AI Act's Annex III high-risk obligations (delayed to Dec 2, 2027) all target AI that makes or materially influences decisions about lending, housing, employment, healthcare or education.
Mitigation: Keep agents to intake, routing and drafting for anything in those categories; require a documented human decision-maker for the consequential step itself.
EU AI Act Article 50 transparency
Applies from August 2, 2026 to any agent talking to people in the EU, including internal tools used by EU-based employees.
Mitigation: Build an AI disclosure into the first interaction for any EU-facing or EU-staffed deployment.
Excessive agency and governance (OWASP Agentic AI ASI02, ASI03)
An agent with broad write access to CRM, ERP or identity systems is the scenario Forrester flags as likely to cause a publicly disclosed, agent-driven breach in 2026.
Mitigation: Scope every agent's tool access to the minimum it needs, require human approval on irreversible actions, and log every tool call.
Financial-services specific rules (FINRA, SEC)
FINRA's 2026 Regulatory Oversight Report has a standalone section on AI agents covering governance, testing, prompt and output logging, and limiting agent system access; the SEC has also charged firms for overstating their AI capabilities ('AI washing').
Mitigation: Archive every agent communication under existing recordkeeping rules and never market agent capabilities you haven't tested.
Vendor contract and data-training risk
Enterprise platform vendors (Agentforce, Sierra, Decagon) process your internal and customer data to run the agent, and contract terms on training and retention vary widely.
Mitigation: Negotiate a no-training clause and data-retention limits before signing an enterprise contract, not after.
Build vs buy: the real math
4-8 weeks for a RapidDev-scoped pilot; 2-6 weeks to a first Copilot Studio agent with an internal team
Custom build time
$13K-$35K for a scoped agency pilot on one process; $0-$30K in engineering time for a DIY build; enterprise platform contracts run $50K-$350K+/yr
One-time investment
A scoped pilot on one process typically breaks even against a six-figure platform contract within the first year, because you're paying for one process instead of a company-wide license
Breakeven vs buying
Sierra and Decagon-style enterprise platforms are reported to start around $50,000-$150,000 a year before you've proven a single use case. A RapidDev-scoped pilot on one high-value process costs $13,000-$35,000 upfront plus $500-$3,000/mo in model and infra costs, meaning you can validate the business case for roughly a tenth of a platform contract's first-year cost. The math shifts toward a full platform once you've validated 3-4 processes and need shared governance, observability and a single pane of glass across all of them, at which point a platform's annual fee starts to beat stitching together multiple point solutions.
Skip the DIY — RapidDev builds the production version
A Lovable MVP gets you a demo. Production needs auth that doesn't leak data, AI calls that don't bankrupt you, observability when models drift, and code you can audit. That's what we ship through our AI app development service.
Discovery call (free)
30 minWe map your exact company-wide AI agent use case: who uses it, target volume, AI model choice, integrations, compliance scope. You get a detailed scope document and fixed-price quote within 48 hours.
AI-accelerated build
4-8 weeks for a RapidDev-scoped pilot; 2-6 weeks to a first Copilot Studio agent with an internal teamOur engineers use Claude Code, Lovable, and custom tooling to ship 3–5x faster than agencies. You see weekly progress in a staging environment — not a black box.
Launch + handoff
1 weekWe deploy to your infrastructure, transfer the GitHub repo, set up CI/CD and monitoring, and train your team. You own 100% of the source code, prompts, and model configurations.
What you get
Timeline
4-8 weeks for a RapidDev-scoped pilot; 2-6 weeks to a first Copilot Studio agent with an internal team
Investment
$13K-$35K for a scoped agency pilot on one process; $0-$30K in engineering time for a DIY build; enterprise platform contracts run $50K-$350K+/yr
vs SaaS
ROI in A scoped pilot on one process typically breaks even against a six-figure platform contract within the first year, because you're paying for one process instead of a company-wide license
30-min call. Fixed-price quote within 48 hours. No commitment.
Frequently asked questions
How much does an AI agent for enterprise cost?
Enterprise SaaS platforms range from $2 per conversation (Salesforce Agentforce) to reported six-figure annual contracts (Sierra, Decagon). A scoped RapidDev pilot on one process runs $13K-$35K upfront plus $500-$3,000/mo in model and infra costs. A DIY build on Microsoft Copilot Studio or AWS Bedrock AgentCore costs $200-$1,000+/mo in credits plus your own engineering time.
How long does it take to set up?
A scoped agency pilot on one process takes 4-8 weeks. An enterprise SaaS platform takes 4-12 weeks depending on how many systems it needs to connect to, plus procurement and security review. An internal build on Copilot Studio or AgentCore can produce a first working agent in 2-6 weeks, with governance maturity taking months longer.
Can RapidDev build this for us?
Yes. RapidDev has shipped 600+ apps and offers a free 30-minute consultation to scope which single process to pilot first, before you commit to a company-wide platform.
Should we pilot one process or roll out a platform company-wide?
Pilot one process first. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027 over unclear value and weak risk controls, and a scoped pilot costs a fraction of a company-wide platform commitment while you prove the business case.
Which cloud platform should we build on?
Whichever one already hosts your systems of record. Microsoft Foundry Agent Service fits Azure/M365 shops, AWS Bedrock AgentCore is framework-agnostic for AWS-standardized teams, and Google's Gemini Enterprise Agent Platform fits Google Cloud shops; switching clouds just to chase a marginally better agent runtime rarely pays off.
What's the biggest governance risk?
Excessive agency: giving an agent broad write access to a production CRM, ERP or identity system without a human approval gate. Forrester predicts an agentic AI deployment will cause a publicly disclosed data breach in 2026, and OWASP's Agentic AI Top 10 lists this among its highest-severity categories.
Can it work with Salesforce, ServiceNow or SAP?
Yes. Salesforce Agentforce and ServiceNow's AI Agents (Now Assist) are native to those platforms, and a custom build can connect to SAP, Oracle or any other system of record through API tool calls, though that integration work is exactly what a scoped agency build is for.
Want the production version?
- Delivered in 4-8 weeks for a RapidDev-scoped pilot; 2-6 weeks to a first Copilot Studio agent with an internal team
- You own 100% of the code
- AI cost monitoring built in
30-min call. No commitment.
