Skip to content
Connector Scout
–saved records

Original research · 25 sources reviewed

The Real ROI of AI Connectors: What Enterprise Evidence Actually Shows

A stand-alone model can improve an answer. The available enterprise evidence suggests the larger financial effect comes from connecting that model to the systems where the work actually happens — but the published ROI numbers need careful reading before they are quoted.

Written and researched by Christian Stewart, Founder and EditorPublished 5 Aug 2026Updated 5 Aug 2026Status: evidence reviewed
How a connector extends a language model into enterprise systemsA central language model node is linked to four enterprise system nodes — knowledge, CRM, support and operations. A band beneath the diagram shows the progression from answering a question, to retrieving current private context, to acting inside the system that owns the work.KnowledgeDocs, wikis, ticketsCRMAccounts, pipelineSupportCases, chatOperationsApprovals, provisioningLLMmodel layerAnswerRetrieveAct
Connector Scout illustration. A model answers; a read connector retrieves current private context; an action connector performs work in the system that owns the process.

What this research is, and what it is not

This is Connector Scout’s original synthesis of public secondary research. It is not a set of proprietary customer interviews, and it is not a randomized experiment comparing the same model with and without connectors. Every figure below is labelled with the kind of evidence behind it.

Research snapshot

The 3.5 million figure combines one reported total with three Connector Scout annualizations of separately disclosed rates. It is evidence of scale, not an average, a benchmark, or a like-for-like comparison.

A stand-alone large language model can draft an email, summarize supplied text, brainstorm ideas, or help analyze information pasted into a conversation.

What it usually cannot do is answer a company-specific question using current internal information. It cannot check the status of a customer renewal, find the latest policy in SharePoint, update a Salesforce opportunity, provision software access, or resolve a support request unless it can reach the systems where that work lives.

That is the economic argument for AI connectors. Connectors move AI from generating an output to participating in a workflow. The available evidence suggests that this distinction matters — but the ROI claims require careful interpretation.

Key findings

Evidence quality key

Evidence quality key

Every figure on this page carries one of these four labels. They are not interchangeable.

  • Stronger evidenceControlled studiesRandomised or otherwise controlled productivity experiments with a comparison group and independent evaluation of output.
  • Directional evidenceInternal surveys and named customer reportsCompany-reported results and vendor-published customer stories. Self-reported, selected for strong outcomes, and not independently audited.
  • Modeled evidenceCommissioned composite ROI analysesVendor-commissioned economic-impact models built on a composite organisation, with assumptions about adoption, labour cost and recaptured time.
  • Illustrative evidenceConnector Scout calculationsArithmetic Connector Scout performed on figures a vendor or company disclosed, or on the reader's own inputs. Not a measurement of any real deployment.

Methodology

Connector Scout reviewed 25 public sources:

  • 18 disclosures from named enterprise deployments
  • Three vendor-commissioned economic-impact studies
  • Four controlled or randomized productivity studies

Those 25 sources are published across the 19 primary URLs listed in the sources section; several of them — including the Moveworks customer index and the Microsoft customer-story round-ups — document more than one named deployment.

The named-company sample covered knowledge search, productivity, customer service, employee support, software development, security, healthcare, legal work, finance, procurement and supply-chain operations.

We included systems that could retrieve private enterprise information or perform work in another application, even when the vendor used terms such as integration, copilot, agent, enterprise search platform, retrieval-augmented generation system or AI assistant rather than connector.

Selection and modelling bias is unavoidable here

Enterprise case studies are normally published because they produced strong results, and commissioned ROI studies rely on assumptions about adoption, labour cost, time recapture and future benefit. Vendor-reported time savings in this article are self-reported; none of them are independently audited. Read them as directional evidence, not as measured averages.

Connector Scout applies the same sourcing standard here as it does to every connector record. See Connector Scout’s verification methodology for the source hierarchy and dating rules.

The AI value ladder

Each rung adds a different kind of value — and a different kind of risk.

  1. Stage 1Stand-alone LLMDraft, summarize, explainPrimary value: Reduces the labour needed to produce a suitable piece of work from supplied context.Biggest limitation: Cannot see current private company data and cannot change anything downstream.
  2. Stage 2Read connectorsRetrieve current private contextPrimary value: Grounds answers in the company's own documents, tickets, records and messages.Biggest limitation: Still ends at an answer; a person performs the remaining work.
  3. Stage 3Action connectorsCreate, update, send, approvePrimary value: Removes handoffs by completing steps inside the system that owns the process.Biggest limitation: Write access raises the cost of an error, so it needs scopes, approval gates and audit logs.
  4. Stage 4Orchestrated workflowsComplete multi-step work and escalate exceptionsPrimary value: Turns a process into something measurable: volume handled, completion rate, escalation rate.Biggest limitation: Requires workflow redesign, reliable source data and a named process owner.

What ROI looks like with an LLM alone

The strongest evidence for stand-alone LLM productivity comes from bounded tasks.

In an MIT experiment involving 453 college-educated professionals, participants using ChatGPT completed occupational writing tasks 40% faster, while independent evaluators rated their output 18% higher. The assignments did not require detailed knowledge of a specific company, customer or factual environment. [MIT News: Study finds ChatGPT boosts worker productivity for some writing tasks (opens in a new tab)]

A randomized field experiment involving 758 BCG consultants found that GPT-4 users completed 12.2% more subtasks and worked about 25% faster on tasks inside the model’s capability frontier. Human-rated quality also increased. Performance deteriorated when participants used the model for a task outside that frontier. [Harvard Business School AI Institute: Navigating the Jagged Technological Frontier (opens in a new tab)] [Harvard Business School AI Institute: Back to the Beginnings of AI at Work (opens in a new tab)]

The boundary of the conversation

A stand-alone LLM can reduce the labour required to produce a suitable piece of work, but that does not automatically reduce the time required to locate the inputs, verify the answer, transfer it into another system, obtain approval, or complete the surrounding process.

This is why individual productivity can rise without producing obvious enterprise-level financial impact. If the saved minutes are absorbed back into the working day, or the bottleneck sits in an approval queue rather than in drafting, the measured gain never reaches a cost line. McKinsey’s work on where AI creates value makes a similar point: impact concentrates where the work itself is redesigned, not where a tool is simply issued. [McKinsey & Company: Where AI will create value — and where it won’t (opens in a new tab)]

Where connectors create additional value

Enterprise knowledge search

Zillow connects more than 30 data sources and 138 million documents through Glean. Internal survey data found 1.5 hours saved per employee per week on information search, with 80% regular adoption across roughly 7,000 employees, and more than 3,400 specialized agents created. [Glean: Zillow customer story (opens in a new tab)]

GCash reports two to three hours saved per employee each week, with adoption above 90% in some departments. [Glean: GCash customer story (opens in a new tab)]

This is the clearest structural argument for connectors: private, distributed, frequently changing data cannot be reliably reproduced by a stand-alone model. No amount of model improvement tells an employee which version of a policy is current in their own document store.

General workplace productivity

Microsoft’s commissioned Forrester study modeled nine hours saved per user each month, 116% three-year ROI and a 10-month payback for a composite enterprise. It drew on interviews with 16 decision-makers across 12 organisations and a survey of 367 respondents. [Forrester / Microsoft: The Total Economic Impact of Microsoft 365 Copilot (opens in a new tab)]

These are mostly time-savings reports rather than financial results. They become financial value only when the time is redirected into output, service levels, avoided hiring or another business outcome someone is accountable for.

Employee support and internal operations

ServiceNow estimates that 89% of customer self-service requests are supported by AI, with a 9.0 CSAT and more than 2.3 million employee hours saved in 2025. Its agents search trusted knowledge, interpret intent, trigger workflows, complete tasks and escalate exceptions under permissions. [ServiceNow: Now on Now: customer self-service (opens in a new tab)]

Moveworks customer examples show the same pattern at smaller scale: West Monroe reported more than 8,000 tickets resolved and another 4,000 accelerated, 40% lower ticket-service costs and about $1.4 million in one year; Procore reported nearly 4,000 support-operator hours saved per quarter; Amadeus reported more than 16,000 employee hours saved per month; ManTech reported about 1,000 annual hours saved on employment verification. [Moveworks: Moveworks customer stories (opens in a new tab)]

Moveworks’ commissioned Forrester model puts a composite 30,000-employee organisation at 256% three-year ROI with 90,000 productive hours reclaimed annually. [Moveworks: The Moveworks Total Economic Impact study by Forrester (opens in a new tab)]

Customer service

Klarna said that during its first month its AI assistant handled two-thirds of customer-service chats, performed work equivalent to 700 full-time agents, reduced average resolution time from 11 minutes to two, lowered repeat inquiries by 25% and was expected to contribute $40 million in 2024 profit improvement. [Klarna: Klarna AI assistant handles two-thirds of customer service chats in its first month (opens in a new tab)]

Customer service illustrates why connector depth matters. Account context decides whether an answer is relevant, company policy decides whether it is allowed, an action decides whether the case is actually closed, and exception routing decides what happens when the model should not proceed. Each of those four capabilities expands the business case, and each one also expands the permission surface you have to review.

Specialized and high-value workflows

A use case does not need to save a large share of company-wide labour to be worth building. It needs to remove enough cost, delay, leakage or error from a process the business already values.

Connector Scout calculation: 3.5M+ annualized hours

One reported total plus three annualizations of separately disclosed rates.

Four deployments exceed 3.5 million annualized hours

Different products, populations, reporting periods and measurement methods. Do not interpret this as a benchmark. Each bar links to the disclosure it comes from; three of the four values are Connector Scout arithmetic on disclosed rates.

Plotted figures and their sources: ServiceNow source: ServiceNow (opens in a new tab); Toshiba source: Microsoft (opens in a new tab) (Connector Scout arithmetic); Zillow source: Glean (opens in a new tab) (Connector Scout arithmetic); Amadeus source: Moveworks (opens in a new tab) (Connector Scout arithmetic)

Annualized hours: ServiceNow 2,300,000 reported for 2025; Toshiba 672,000; Zillow 403,200; Amadeus 192,000. The last three are Connector Scout annualizations of disclosed rates.

Annualized hours saved across four disclosed deployments, with the methodology for each figure
DeploymentAnnualized hoursHow the figure was producedSource
ServiceNow source: ServiceNow (opens in a new tab)2,300,000Reported by ServiceNow for 2025 (employee hours saved).ServiceNow: Now on Now: customer self-service (opens in a new tab)
Toshiba source: Microsoft (opens in a new tab)672,000Connector Scout annualization: 5.6 hours/month × 10,000 employees × 12.Arithmetic by Connector Scout on the disclosed rate.Microsoft: AI-powered success with 1,000 stories of customer transformation and innovation (opens in a new tab)
Zillow source: Glean (opens in a new tab)403,200Connector Scout annualization: 1.5 hours/week × 48 weeks × 7,000 employees × 80% adoption.Arithmetic by Connector Scout on the disclosed rate.Glean: Zillow customer story (opens in a new tab)
Amadeus source: Moveworks (opens in a new tab)192,000Connector Scout annualization: 16,000 hours/month × 12.Arithmetic by Connector Scout on the disclosed rate.Moveworks: Moveworks customer stories (opens in a new tab)

Combined, those four deployments represent 3,567,200 annualized hours. The Toshiba, Zillow and Amadeus values are Connector Scout calculations on separately disclosed rates; only the ServiceNow total is reported as an annual figure. All four use different products, populations, reporting periods and measurement methods, so treat this as evidence of scale, not as an average or a benchmark. [ServiceNow: Now on Now: customer self-service (opens in a new tab)] [Microsoft: AI-powered success with 1,000 stories of customer transformation and innovation (opens in a new tab)] [Glean: Zillow customer story (opens in a new tab)] [Moveworks: Moveworks customer stories (opens in a new tab)]

Connectors versus an LLM alone

The same model produces different economics depending on what it can reach.

Capability, value source and typical limitation at each level of connection
ConfigurationWhat it can doPrimary source of valueTypical limitation
Stand-alone LLMDrafts, summarises, explains, and analyses text a person supplies in the conversation.Faster production of a suitable individual output.No current private context, and no change in any system of record.
LLM with read connectorsSearches and retrieves current documents, tickets, records and messages the account is allowed to see.Time removed from finding, assembling and verifying inputs.Answer quality depends on indexing, permissions and data hygiene; the work still ends with a person.
LLM with action connectorsCreates and updates records, sends messages, files requests and submits approvals.Removed handoffs and tool switching inside a defined process.Errors now have consequences, so scopes, approvals, logging and reversibility matter.
Orchestrated AI workflowCompletes multi-step work end to end under permissions and escalates exceptions to a human.Measurable throughput on a high-volume process, not just individual assistance.Needs workflow redesign, reliable data, monitoring and a clear process owner.

Which of these you can actually buy depends on the platform. ChatGPT, Claude and Gemini connector coverage compared sets out where read-only access ends and action support begins, and the connector relationship directory holds the underlying documentation-verified records.

What does ROI generally look like?

Observed outcomes by scenario, with the confidence Connector Scout places in each
ScenarioWhat the evidence showsConfidenceEvidence type
Stand-alone LLM on suitable tasksAbout 25%–40% faster task completion; rated quality may also improve.Sources: MIT News (opens in a new tab), Harvard Business School AI Institute (opens in a new tab)Relatively strong for narrow controlled tasks.Stronger evidence
Connected knowledge and productivity toolsRoughly one to five hours saved per active user per week in several named deployments.Sources: Glean (opens in a new tab), Glean (opens in a new tab), Microsoft (opens in a new tab)Moderate; mostly internal surveys.Directional evidence
Connected support workflowsMaterial ticket deflection, faster resolution and lower support labour in well-defined cases.Sources: ServiceNow (opens in a new tab), Moveworks (opens in a new tab), Salesforce (opens in a new tab)Moderate; metrics are inconsistent between companies.Directional evidence
Enterprise-wide financial return116%–256% modeled three-year ROI across three commissioned Forrester studies.Sources: Forrester / Microsoft (opens in a new tab), Glean (opens in a new tab), Moveworks (opens in a new tab)Directional only.Modeled evidence
High-volume operational workflowsSeven-figure annual impact is possible when the workflow reduces leakage, external cost or large volumes of manual work.Sources: Klarna (opens in a new tab), Microsoft (opens in a new tab)Highly case-specific.Directional evidence

Modeled three-year ROI in commissioned Forrester studies

Vendor-commissioned composite economic models; not an independent industry average. Each modeled figure links to its Forrester Total Economic Impact study below.

Plotted figures and their sources: Microsoft 365 Copilot source: Forrester / Microsoft (opens in a new tab); Glean source: Glean (opens in a new tab); Moveworks source: Moveworks (opens in a new tab)

Modeled three-year ROI: Microsoft 365 Copilot 116 percent, Glean 141 percent, Moveworks 256 percent.

Modeled three-year ROI in three vendor-commissioned Forrester Total Economic Impact studies
DeploymentModeled three-year ROIHow the figure was producedSource
Microsoft 365 Copilot source: Forrester / Microsoft (opens in a new tab)116%Composite enterprise; 10-month payback; 50% of saved time modeled as recaptured.Forrester / Microsoft: The Total Economic Impact of Microsoft 365 Copilot (opens in a new tab)
Glean source: Glean (opens in a new tab)141%Composite organisation using Glean enterprise search.Glean: Forrester Study: The Total Economic Impact of Glean (opens in a new tab)
Moveworks source: Moveworks (opens in a new tab)256%Composite 30,000-employee organisation; 90,000 productive hours reclaimed annually.Moveworks: The Moveworks Total Economic Impact study by Forrester (opens in a new tab)

Why connectors can produce more value

Remove information gathering

Search, retrieval and assembly are pure overhead. A read connector removes them from the task rather than making them faster.

Reduce tool switching and handoffs

Every switch between systems is a chance to lose context, wait for someone, or transcribe something incorrectly.

Make repeatable, measurable workflows possible

Once a process runs through a defined path you can count volume, completion, escalation and reversal — which is what a finance team needs.

Compound value with volume

A small saving on a rare task is noise. The same saving on a high-frequency transaction becomes a line item.

High frequency × meaningful labour per transaction × reliable automation × strong adoption

Weaken any one term and the return falls disproportionately.

A practical ROI formula

Annual productivity value = hours saved per transaction × annual transaction volume × fully loaded hourly cost × time-realization rate

Then add, where you can defend the number: avoided hiring or outsourcing, reduced software and service cost, lower error, rework and leakage, incremental gross profit, and quantified risk reduction.

Annual cost should include licensing, model usage, implementation, data cleanup, security and governance work, training, monitoring, evaluation and maintenance.

ROI = (annual benefit − annual cost) ÷ annual cost

Time realization is the term most business cases get wrong. Forrester’s Microsoft model assumed only 50% of saved time was recaptured as productive value; assuming 100% roughly doubles the modelled benefit without any change in the deployment. [Forrester / Microsoft: The Total Economic Impact of Microsoft 365 Copilot (opens in a new tab)]

At the calculator defaults below — $900 annual cost per user, $75 fully loaded hourly cost, 50% realization and 48 working weeks — saving roughly 30 minutes per user per week is the break-even point. That break-even is Connector Scout’s own arithmetic on those inputs, not a published result; only the 50% realization rate is taken from a source (Forrester / Microsoft: The Total Economic Impact of Microsoft 365 Copilot (opens in a new tab)), and the cost and labour-rate figures are illustrative placeholders. Above that, the programme returns money; below it, the licence is a cost centre no matter how positive the user feedback is.

Illustrative ROI calculator

Connector ROI calculator

Illustrative calculator — not a forecast. Everything is computed in your browser and nothing is sent anywhere.

Illustrative
Annual productivity benefit
$90,000
Annual cost
$90,000
Net annual benefit
$0
ROI
0%
Estimated payback
No payback at these inputs

All arithmetic below is Connector Scout’s, not a published study result. Benefit = users × (minutes saved ÷ 60) × working weeks × hourly cost × realization rate. Cost = users × annual cost per user. ROI = (benefit − cost) ÷ cost. Payback = cost ÷ benefit × 12, shown only when the annual benefit exceeds the annual cost.

How to prove ROI

  1. 1Baseline the process before enabling AIWithout a pre-AI baseline there is nothing to compare a saved hour against.
    • Search and information-gathering time
    • Cycle time and labour per transaction
    • Volume, escalations and error rate
    • Satisfaction and revenue or cost leakage
  2. 2Distinguish assistance from completionPrompt counts are activity. Completion is the thing that changes a cost line.
    • Answers generated and answers accepted
    • Tasks initiated and tasks completed
    • Tasks completed without human intervention
    • Escalated, reversed or corrected outcomes
  3. 3Use a comparison group or phased rolloutA held-back team or a staged launch separates the connector's effect from seasonality, staffing changes and process improvements happening anyway.
    • Matched control team or region
    • Staggered enablement dates
    • Same metric definitions on both sides
  4. 4Measure realized value, not theoretical timeSaved minutes become financial value only when they are redirected into output, service, avoided hiring or reduced spend.
    • Volume handled per person
    • Avoided hiring or outsourcing
    • Reduced software, service or vendor expense
    • Cycle time, quality, revenue, burnout and turnover
  5. 5Audit permissions and the cost of failureA connector's downside is asymmetric: a wrong answer is recoverable, a wrong write may not be.
    • Account scopes and what the indexed corpus exposes
    • Retention, audit logs and revocation paths
    • Approval gates before writes
    • Consequence of an incorrect action

Step five is where most programmes underestimate their exposure. Work through permissions and data access before enabling writes, and confirm what you are actually deploying by reading connectors versus MCP servers — the ownership and revocation story differs sharply between the two.

Where enterprises should start

Choose a workflow that is frequent, valuable, repetitive but not completely rigid, supported by reliable data, measurable, reversible when it goes wrong, and owned by a clear process owner. If any one of those is missing, fix it before buying licences.

Candidates that recur throughout the evidence: knowledge search, employee support, customer-service triage, meeting follow-up, sales preparation, document review and structured reporting. Role-level starting points are mapped in the data analyst connector guide and the digital marketer connector guide.

Start with read access. Add constrained action permissions only after answer quality, adoption, failure handling and escalation paths are proven. Shortlisting is easier in the side-by-side record comparison.

Bottom line

Stand-alone LLMs create legitimate value, and controlled studies show roughly 25% to 40% faster work on suitable tasks. But that value often stops at the boundary of the conversation. [MIT News: Study finds ChatGPT boosts worker productivity for some writing tasks (opens in a new tab)] [Harvard Business School AI Institute: Navigating the Jagged Technological Frontier (opens in a new tab)]

Connectors extend the model into the operating environment: current context, search, handoffs and — when action access is enabled — parts of the actual process.

The honest conclusion is not that every connector produces triple-digit ROI. It is that AI creates the clearest financial return when it is attached to a measurable workflow, trusted data, and an outcome the business already values.

“A better model can improve an answer. A well-designed connector can remove the work surrounding it.”

Frequently asked questions

Sources

Grouped by evidence type. Vendor customer stories and commissioned models are vendor-reported and not independently audited.

Stronger evidence: controlled studies

Randomised or otherwise controlled productivity experiments with a comparison group and independent evaluation of output.

Directional evidence: internal surveys and named customer reports

Company-reported results and vendor-published customer stories. Self-reported, selected for strong outcomes, and not independently audited.

Modeled evidence: commissioned composite roi analyses

Vendor-commissioned economic-impact models built on a composite organisation, with assumptions about adoption, labour cost and recaptured time.

Related research